Chat:World/2021-05-21
NotSureWhyThisWorks: when you have the exact same algorithm as your opponent. https://www.codingame.com/replay/558868531
NotSureWhyThisWorks: another one: https://www.codingame.com/replay/558869098
pmor: lol
derjack: good morning
VizGhar: morning sunshine
derjack: oO
pshemke: D:
KiwiTae: :') NguyenThanhTam wrong website
ArtLiteracyStory: Anyone want to play with a streamer together? https://www.twitch.tv/senamonry is playing
**BrainNotFoundException slaps everyoe around a bit with a large fishbot
**BrainNotFoundException slaps everyone around a bit with a large fishbot
Cdude: hi
Cdude: Container Terminal
Habiasa: Error loading certifications
Habiasa: What does it means?
KiwiTae: refresh page
CortexDeveloper: Hi Can you recommend books or useful link to learn AI? thnx
Cdude: https://www.codingame.com/multiplayer/bot-programming/coders-strike-back
VizGhar: Compilation took too long and has been interrupted... o.O
KiwiTae: >< linux kernel?
VizGhar: codingame IDE :)
KiwiTae: ouch
VizGhar: they are probably updating kotlin or something
jacek: good old friday deployment
VizGhar: plain code is compiled in time (however it still takes way too long)
Tychkorg69: stfu yall work at callcenters
Tychkorg69: while i work at google
Greg_3141: since you work at google, could you google who asked?
AntiSquid: CortexDeveloper https://nnfs.io/ i recommended it to many people, including users here on CG and I've seen only positive responses so far
CortexDeveloper: :thumbsup:
yassine_Benlahbib: hello
jacek: so youre reinforced on positive values only
Alina24: hi
MSmits: AntiSquid I am reading that nnfs book
Wontonimo: how is it? I have watched a few of the authors videos back in the day before he wrote the book
MSmits: it's too bad the video series only goes to 25% of it. I actually thought it was better than the book. Normally I prefer written stuff
MSmits: but the book is fine
MSmits: I'm at around page 200
MSmits: I know all of the math so every time it's about math, I speed through ot
jacek: any plot twists?
MSmits: it
MSmits: it's just basic NN stuff. Pretty in depth. doesn't get into convolution and such, but I does explain everything up to and including backpropagation
MSmits: explains many activation functions too
MSmits: and, as opposed to other sources, explains every step of the math in detail
MSmits: goes into saving and loading of models and such at the end
MSmits: and it's not written with tensorflow or other tools in mind. Just basic python
MSmits: well, numpy of course
MSmits: but it even explains the workings of numpy
KalamariKing: nnfs yes?
MSmits: oh and there's lots of code. You can reproduce everything they do
MSmits: yes
KalamariKing: great resource
Wontonimo: nice. seems like a really really long book for just fully connected linear NN
KalamariKing: Yeah kinda
MSmits: it's long because of detail
MSmits: if you go through it, you don't just know how to do things. You know how it works and why it works
MSmits: and it's not long at all if you have prior knowledge. I am a third of the way through in about a day
MSmits: the information contained in here, is enough for a university class
MSmits: which would take 8 weeks or something
MSmits: but gj AntiSquid for sharing. Darkhorse told me about it, but I guess you told him?
KalamariKing: It goes pretty far in-depth, not sure about other types of nns tho
Wontonimo: On the topic of NN (and AI in general), This MIT OpenWare course is really good survey of AI https://www.youtube.com/watch?v=uXt8qF2Zzfo
Wontonimo: the link is to the NN video
MSmits: yeah i guess it covers similar stuff, but the nnfs has way more detail than a single video can
MSmits: (unless it's 10 hrs)
MSmits: another major advantage of this book is that because it contains actual code, it really invites you to try stuff
MSmits: I think a lot of people watch the video's and such but never actually start experimenting
Wontonimo: what? how do you learn then?
MSmits: you dont learn how to do it, you learn about it. I think many people just learn about it
MSmits: like i have for the past year o rso
Wontonimo: ;)
KalamariKing: A lot of people I know who supposedly know nn stuff just know theory
KalamariKing: kinda sad
MSmits: as long as enough know the practice, it's fine
MSmits: I wish more people who didn't do science, knew about science
MSmits: then we might not have people lighting 5G towers on fire
MSmits: they dont have to ever go into a lab to prevent that you know
MSmits: they just have to know the theory
jacek: i know nn by practice, little on theory. am i bad
jacek: it works and i dont know why
MSmits: you know a lot of theory
MSmits: you just dont realize :P
MSmits: it's imposter syndrome
MSmits: what do you not know?
MSmits: I mean, you say you dont know why stuff works, why not?
jacek: https://xkcd.com/1954/
jacek: i was a little sarcastic
MSmits: lol you have a meme or link for everything :)
jacek: i just didnt dive into math too deep
jacek: there is xkcd for everything
MSmits: well I'm sure you know what a partial derivative is, or ?
jacek: yes
MSmits: then you're fine
MSmits: backpropagation is the most complicated part and thats just a whole bunch of partial derivatives
MSmits: i dont think you need to go through the math more than 1 time. Consider why some things work and some things dont. Then abstract it and dont look at it again unless you have to
MSmits: it's helpful to know why you would use a softmax for the last activation function
MSmits: that sort of thing, but i guess you know that too
Wontonimo: like all of programming, solve the thing, make it into a function, then use it at a high level and move on
MSmits: yeah
MSmits: oh i had a question jacek
MSmits: maybe you know this
MSmits: when you take the input of a typical board game, which has cells that generally have p1, p2 marks or are empty
MSmits: is it useful to use as input -1 for p2, 0 for empty and 1 for p1
MSmits: or do you have 2 inputs per cell that are 0 or 1, 1 for each player
MSmits: the first version is way more compact, but i suspect there are disadvantages to it
Wontonimo: i don't know about optimizing for CG, but in image processing you want separate layers
MSmits: layers of what?
jacek: for me each state is separate input
KalamariKing: Wontonimo is talking nns
Wontonimo: 2 inputs
jacek: 0 activates different input, X activates different input and O activates different input
KalamariKing: If you're doing a nn, I would one-hot the inputs
jacek: ^
MSmits: but doesnt that mean 3 inputs?
KalamariKing: Or two-hot or whatever, just not smth like -1 for a, 0 for b, 1 for c, etc
KalamariKing: Why does the number of inputs matter?
jacek: well yeah, for TT that would be 3 inputs x 9 so 27 inputs in which there are always 9 activated
MSmits: ok so thats what you do, that's actual one-hot indeed
MSmits: i am just wondering *why* the -1,0,1 version wouldnt work
MSmits: or maybe if it works, but far less efficiently
KalamariKing: The nn nodes are linked together such that a -1 might give the opposite result than a 1, if that makes sense
MSmits: sure
KalamariKing: If its one-hot then they can give similar but different results
MSmits: but is that a bad thing for a board game?
MSmits: an opposite result?
Wontonimo: I need to do a multi with NNs before I say anymore. This format of super tight compute time, small(ish) memory constraints, and no libraries is out of my comfort zone. Perhaps the -1/0/1 is better for CG constraints.
KalamariKing: That is also something to keep track of
KalamariKing: But couldn't you cut time somewhere else? 100ms is a lot of time to a computer
MSmits: I was thinking about oware. If you input the number of seeds, you can have like 60 different possible values for the input. That's way more even. If you have only 3 for a cell, why would that be an issue if 60 in oware isnt
MSmits: the seeds in oware are also discrete possibillities
jacek: that -1/0/1 thing maybe would work if you flip the board for the current player
Wontonimo: okay MSmits, consider the case where the NN would like to calculate the XOR between two adjacent cells and check if only one is empty.
MSmits: i'm thinking
MSmits: why would this be a problem?
jacek: fyi i have 820 inputs for oware
Wontonimo: i'm thinking also...
MSmits: brb, keep talking:)
KalamariKing: Was that like a rhetorical question or an actual question
Wontonimo: sorry, should have stated that it's what i'm starting to think about as a mental test
Wontonimo: i don't have the answer yet
jacek: what was the question o.O
Wontonimo: given 2 cells in a board,
Wontonimo: and it can be owned by p1, p2 or unoccupied
KalamariKing: if the nn wanted to xor two adjacent cells, to check if only one is empty
Wontonimo: yes. using either -1/0/1 for encoding or 1 hot
KalamariKing: would one-hot x3 inputs be better or a single [ -1/0/1 ]
Wontonimo: 1 hot being [0,0,1] [0,1,0] [1,0,0]
KalamariKing: yeah, where only... one is... hot
jacek: why 3 brackets?
KalamariKing: Three test input cases
Wontonimo: two -1/0/1 (one for each cell) or two [x,x,x], one for each cell
jacek: 2 vs 6 inputs
Wontonimo: right
KalamariKing: One-hot simply means instead of giving the nn one input ranging from x to y, you give it (y-x) inputs and only have one on
KalamariKing: if that makes sense at all
jacek: well both would work i think
jacek: when you have bias input anyway
Wontonimo: float x[2] vs float x[2][3]
MSmits: jacek holy sht so many inputs for oware
jacek: as long as it works...
MSmits: yeah, it just seems a lot
MSmits: this is basically one-hot for seeds then?
MSmits: + score
jacek: 2 * 26 for scores, 12 * 32 * 2 for houses
jacek: probably i could get away with normal inputs for scors
MSmits: why the *2 for houses?
MSmits: or is it 6 * 2 ?
jacek: 0..31 seeds, x2 for side to move
MSmits: oh ok, you could flip there
MSmits: symmetry
jacek: mhm
MSmits: player to move is always lower houses
MSmits: flipping is super chea[
jacek: ah well
MSmits: cheap, negligble
jacek: welp robo does the flip
darkhorse64: my nnfs book recommandation is independant from Antisquid's. It looks like we are at least 3 to see its value for nn noobs
MSmits: it's good, because it's so well structured, it's easy to skip the stuff you know
Wontonimo: update, I pulled the -1/0/1 into my xor nn that uses 5 hidden neurons and it can't solve xor
MSmits: i mean i dont mind rereading about weights and biases, but i dont need to learn the derivative of x squaered is 2x :)
MSmits: Wontonimo but xor doesnt use -1 does it ?
RoboStac: for the -1,0,1 thing earlier, my view has always been it doesn't make much difference for the 1/-1 case but it makes the 0 case really hard to deal with (as there isn't a single weight + bias that can make that positive and the other two negative so it would always need extra layers)
AntiSquid: you mean one hot encoding? @_@ why call it one-hot ... like hot what? hotdog ?
MSmits: but would it be bad to have extra layers?
MSmits: one of the possibilities is 1
MSmits: alll others are 0
jacek: more expensive
MSmits: is *hot*
MSmits: 1 is
Wontonimo: conceptually MSmits , can 1 hiddenlayer learn to output true if only one of the two -1/0/1 inputs is 0
Wontonimo: and the answer is , i don't think so
MSmits: it seems like a fun experiment anyways. It's easy to test with a simple TTT example
RoboStac: extra layers are bad if you have processing time / size limits to deal with. Even without it'll probably increase training time with no real benefit
AntiSquid: maybe darkhorse64, i recommended sentdex before he brought up his book though :P hoped he'd stream one of the contests, never happened though
MSmits: RoboStac do you then also have hundreds of inputs for oware?
MSmits: like jacek ?
RoboStac: yeah, 32x14
RoboStac: though I only changed to that a couple of weeks ago
RoboStac: after jacek convinced me
MSmits: what did you have before?
RoboStac: before that it was 14 floats
MSmits: that worked quite well though
jacek: but this one works weller
RoboStac: yeah, but not as well (especially from a training / size perspective)
MSmits: hmm ok, and i guess you clamp the 32, incase the rare 33 seed happens
KalamariKing: what game is this for?
MSmits: oware
darkhorse64: I'll gladly acknowledge you were the first
jacek: my recent come back at oware was made by making the nn predict 6x times faster, and overall 2.5-3x more nodes explored
jacek: and really i dont know how
jacek: my modification should give at most 2x speedup for nn part
RoboStac: I probably should experiment with oware more, this was just the first one trained after changing inputs
MSmits: there could be something that autovectorized
MSmits: jacek
MSmits: and didnt before
MSmits: because you simplified the code
jacek: heck, now even the nn isnt the bottleneck now, but move generations and make/unmake
MSmits: mmh you need my sim :P
darkhorse64: bitboard it !
MSmits: weird though
jacek: i used similar method for uttt and yielded almost 2x iterations speedup
MSmits: i thought sim was usually negligible in performance compared to nn cost
RoboStac: that may be some part of why the one hot one is much better for me - getting rid of the extra layers meant it is now 3x faster
Wontonimo: another update from training... after changing the init values of the NN, it can learn using -1/0/1 if out of 2 cells only one is empty
jacek: yeah, once again i did something that works and i dont know why
MSmits: i guess the large number of inputs does not affect the speed much because it's just the first layer thats more expensive
RoboStac: even though the first layer is so much bigger it's not slower because you still are only actually doing 14 sets of weights
jacek: oh i have it in lookup table, this is really only 14 * HIDDEN for loop
RoboStac: you just skip lots instead
MSmits: ohh right
MSmits: I guess there would come a point where you have to let go of the one hot
MSmits: say the discrete possibilities would be much larger
Wontonimo: here is the tensorflow python code to prove my claim https://gist.github.com/panchishin/a0a6bd3b9119aba6818e95908c54aa49
jacek: like unlimited score
RoboStac: yeah, I was trying to go full one hot for the contest but couldn't make that sensible for suns/score
MSmits: sure yeah
Wontonimo: i *like* the idea of one-hot more though
MSmits: Wontonimo no proof needed, i believe you. Will try stuff myself anyways :)
Wontonimo: I'm sharing. Not about not believing me.
Wontonimo: cool cool
jacek: one hot in oware just grasps the concept of empty house, 1-2 seeds being different than other houses
jacek: some domain knowledge
MSmits: Wontonimo i'll save the code, will get back to it at some point, thanks
KalamariKing: oware is cg right
MSmits: es
KalamariKing: wait its literally mancalla
MSmits: yes
MSmits: no its not
MSmits: mancala is a different game
MSmits: there are many games of this type
MSmits: same board, different rules
KalamariKing: oh I didn't read too much into it, lemme check it out
MSmits: walking the dog for a while, ttyl :)
KalamariKing: have fun
MSmits: thanks
KalamariKing: I heard someone talking about actor-critic for this, how would that work
Wontonimo: I think that was used to mean "value function". If that's so, then you'd use the NN to judge the value of the current board. Good for eval at the bottom of minimax, or other non-terminal search
KalamariKing: Yeah, I get all that, but someone was talking about training an actor-critic, possibly as player and opp respectively, and then uploading the actor
KalamariKing: No search func
MSmits: hmm actor critic is different in that it outputs a move instead of a value using a policy right?
MSmits: or possibly both
AntiSquid: you can design the output independant from the algo itself
AntiSquid: actor critic is almost like having 2 different NN in one
Wontonimo: ah, sorry KalamariKing
KalamariKing: Yeah, and they almost train each other
AntiSquid: you have the learner and the teacher
AntiSquid: or critic
MSmits: isnt azero using this?
MSmits: because it learns both policy and value
MSmits: actor and critic
AntiSquid: no, i think an example would be something like image identification with adverserial attacks such as adding one pixel to the image to confuse the NN whether it's really a bear
KalamariKing: for e.x. generating handwriting, because that's my most recent actor-critic (simple ik, just basics) The critic learns what a '4' looks like, while the actor generates pretty-much-random images The critic can then update the actor; the end result is they both slowly learn to produce handwriting
MSmits: hmm ok
KalamariKing: Kinda interesting, and different than nns like cnns, since its generating the images from noise
AntiSquid: i shared this before, but here it is again, writing A3C from scratch: https://www.youtube.com/watch?v=OcIx_TBu90Q&t=2170s
MSmits: this says that azero also uses actor/critic
MSmits: https://marksaroufim.medium.com/how-to-solve-board-games-8375847c6577
AntiSquid: and before he releases a book, i advertised him first :P
AntiSquid: actually i think he has a book too somewhere
KalamariKing: But how would you apply actor critic to a board game
KalamariKing: Would it look for 'good moves'? or would they progressively train each other?
AntiSquid: i misremember then MSmits, maybe i've seen this info and forgot about it
MSmits: sec KalamariKing i'll explain
AntiSquid: i do too much theory study to be honest
AntiSquid: like anyone heard of transformers here? :P
jacek: the cars?
AntiSquid: no ...
KalamariKing: the voltage booster?
AntiSquid: https://medium.com/inside-machine-learning/what-is-a-transformer-d07dd1fbec04
MSmits: ok so KalamariKing, azero has two things in its mcts search. The network provides a good move (policy) and a value (critic)
AntiSquid: heard about LSTM or GRU ?
MSmits: when you train it, it learns both at once
MSmits: and i guess implicitly the two parts teach eachother
KalamariKing: Oh cool, ok thanks
MSmits: as for our earlier discussion about one-hot and input. You can have a similar discussion about score and W/L output. A game can have an endgame score that decides the winner and you can choose to let the network predict the score or predict the winner
KalamariKing: anti yeah, why
MSmits: in oware, if you precict score instead of winner, you can use scoreless states and not have score-input
MSmits: there are 20 different ways to have a specific pit-layout with 28 seeds on them, with scoreless it is only 1. Wonder if that gains you something
AntiSquid: because those help you understand the theory of transformers better
AntiSquid: KalamariKing
KalamariKing: I see that they're different, but how?
KalamariKing: scalability?
AntiSquid: arhitecture, efficiency ... data required to train, check the article
AntiSquid: amount of data *
AntiSquid: well someone explained it as "it translates stuff into a language of it's own" before translating it into the output ... which is a bit of a round about explanation, but it works
AntiSquid: but then same could be said about any model that has a form of memory
jacek: :thinking:
fenrir: the words actor/critic may be used in different ways, but for AZ I think the right one is policy/value
MSmits: could be, terms are confused a lot in AI
fenrir: where I know acrot/critic is in refinforcement learing (REINFORCE/A2C/A3C) where the actor is the policy, and the critic is used to 'critic' the actual move (good/bad)
fenrir: it is not the actual game value
MSmits: ahh ok
fenrir: (at least while training)
fenrir: it is more like the value the actor willhave if it play every move using its policy
MSmits: I see
7_ajinkya: how're you guys so fast man... unbelievable
jacek: their wife must be complaining tho
MSmits: clashers that get a repeat problem, just paste in their previous answer 7_ajinkya. There are around 600 unique problems
MSmits: it takes a while
MSmits: before they repeat
MSmits: most of us here arent clashers though, including me. Just about everything else on CG is more interesting than coc
7_ajinkya: oh i see... coc is fun though for building quick logic... i anyway code just for fun
MSmits: as do most of us :)
MSmits: here anyways
fenrir: I think in AZ, the value is more the actual game value, the policy being only a 'guide' to avoid wasting time looking for useless node. But well, I think it is heavily biased by the policy
MSmits: yeah thats how the mcts part works, but i wonder if in training the two parts might have an actor/critic kind of relationship. I don't know enough about the subject though
jacek: i think this is more for alphago (before zero). policy and value were separate networks
MSmits: oh right
jacek: mcts search moves by policy and value network tells how good they were
MSmits: policy was trained with supervised learning right?
MSmits: before zero
fenrir: it's more the opposite, you search by value, the policy allows you to look at the most promising node instead of every nodes
MSmits: you search by both afaik
fenrir: in AZ, you may use seperate network or not, that's independant
fenrir: it may help or not
MSmits: separate network helps in situations where you need *just* the policy or *just* the value
MSmits: assuming the two networks would be smaller
jacek: implementation detail :shrug_tone1:
reCurse: Sure but when do you need just one or the other
reCurse: I always miss the interesting convos
jacek: ok lets back to discussing clashes
MSmits: no idea reCurse
RoboStac: whether they're separate or not they're still trained to work together and removing one would most likely make the whole system weaker
MSmits: hey i have an unanswered question for you reCurse.
MSmits: Do you think it would be worth it to make oware predict score instead of win
MSmits: to then use scoreless states and avoid score input
reCurse: What do you expect to achieve with that?
RoboStac: I still don't understand how you think that can work with a tree search
MSmits: i used it before RoboStac
MSmits: well
MSmits: the statespace is about 20x smaller
MSmits: with scoreless states
reCurse: Ok... and?
RoboStac: but if two different sets of moves end up in the same state (but with different scores) they need to eval differently
MSmits: they dojt
MSmits: the best move is the same
MSmits: independent of history
ZoteLeRedoutable: hello i'm making a clash of code game, i haven't find the solution of my own problem x) if u wanna take a look: https://www.codingame.com/contribute/view/677683039c99ce3a75af342d1f0e925a9a26
MSmits: you always want to get the maximum score out of the current state
MSmits: doesnt matter how many points you already have
jacek: i fail to see where would you put that in tree search
RoboStac: thats fine for a single best move from a current position
MSmits: well the value in your mcts, is the predicted future net-score
MSmits: instead of your chance t owin
MSmits: you backpropagate expected seeds gained
RoboStac: but that's not even possible to backprop unless it knows the score
reCurse: I'm still at the expected benefit part
MSmits: RoboStac i mean expected seeds gained in the future
RoboStac: eg after one move you win 5 seeds - every move after that now has a lower score to backprop
MSmits: not the past, it's score
MSmits: less
MSmits: yes, when you backpropagate you add captured seeds
MSmits: so that doesnt matter
jacek: do you mean only your own seeds?
MSmits: it makes backpropagation a bit more complicated
MSmits: no it's a net seed thing
jacek: or your seeds - opponent seeds
MSmits: yes that
MSmits: ido this in my meta mcts
MSmits: it works
MSmits: it doesnt maximize win, it maximizes seeds gained
MSmits: reCurse i am getting to your point
RoboStac: yeah, but that isn't what you want to maximise
RoboStac: you want to maximise win
MSmits: it's easy to combine this with an endgame book
RoboStac: even if it's just by 1 seed
MSmits: because endgame books have seed counts
MSmits: then you can stop at 9 seeds on the board, even at depth 5 in the search or something
MSmits: i know you want to maximize win, but maximizing seeds maximizes win as well
reCurse: Whether it's seeds or wins shouldn't matter
JohnCM: the tree search is a way of policy improvement in az. You use it to improve the move probabilities
MSmits: right
MSmits: so if seeds allows you to use some extra tools like an endgame book at a leaf in your tree, isnt that a benefit?
reCurse: That's a matter of state not evaluation
MSmits: of course
RoboStac: if you've got an endgame book you know the winchance too though
MSmits: but if you predict the number of seeds gained
MSmits: then you can replace the prediction with the book
MSmits: and you have an exact prediction
MSmits: (At some point)
MSmits: at some depth i mean
reCurse: You can do this with a regular win evaluation though
MSmits: yeah, with a large degree of confidence, sure, but not as good as an endgame book will do
reCurse: No
reCurse: When you reach the state with endbook
reCurse: You backprop the book result instead
MSmits: oh yes if the root is there
reCurse: No need to touch evaluation
MSmits: but i am talking about leafs in your searcg
MSmits: leafs can also get to the book
reCurse: Same thing?
MSmits: hmm
RoboStac: if the endgame book says this position is +4 seeds you can work out if you win or lose from that position
MSmits: yes thats true
RoboStac: and just backprop that
MSmits: you're correct, I missed that
MSmits: then the only advantage is scoreless states. Does it help to have a smaller statespace?
MSmits: when solving games it does help a lot and people generally do this
jacek: because it takes less GB in your disk?
MSmits: less solving time too
MSmits: the game simply becomes less complex
reCurse: This is very different from solving
MSmits: i know
MSmits: and i know very little about it in practice, hence my question
reCurse: If the input matters then it will learn to ignore it
reCurse: *doesn't
MSmits: it will of course matter if you dont treat the state as scoreless
MSmits: the only way to make it not matter, if the search handles the state as scoreless
MSmits: (different backpropagation method)
MSmits: but that requires to predict seeds instead of wins
MSmits: you need to know *by how much* you win
reCurse: You're asking about if it's better with less state space
reCurse: I'm telling you it can simplify the state space if it doesn't see it helping
reCurse: The extra info
MSmits: hmm maybe i should rephrase
reCurse: What you output is completely separate
MSmits: will it reduce calculations required
MSmits: if there's less input
reCurse: Very little
MSmits: or will my scheme require a bigger network in the end
MSmits: which means more calculation
reCurse: It will literally only matter on the first layer
reCurse: Then the rest is the same
MSmits: i am not sure, the score is in fact very important if you dont do it scoreless
jacek: but i have only one layer :v
MSmits: so the score is going through the network in an important way
reCurse: Well yeah obviously
reCurse: You're mixing too many topics at once
MSmits: i guess so yes
MSmits: cant help it :)
reCurse: What you output is independent
reCurse: But if you ask it to output win
reCurse: And you withhold critical info
reCurse: Then yes it's bad
MSmits: yeah thats bad
MSmits: i get that
reCurse: So if you don't input score you need to predict seeds gained, naturally
reCurse: Does it help on calculation, barely
CamTheHelpDesk: did rankings not update last night? 🤔
reCurse: You have X * H connections at the first layer
MSmits: maybe on convergence in training?
reCurse: So you brought that down to (X-1) * H
MSmits: no idea how this would work out
reCurse: The other layers are identical
reCurse: You might save some time if that input was useless, it will need to weigh it down first
MSmits: what if you need less nodes in the layer if you dont input score
MSmits: in the 1 hidden layer or 2 or whatever
reCurse: That's a different topic
MSmits: sure
reCurse: Determining the size of what will work
reCurse: Is a whole other can of worms
MSmits: i understand that my question has many layers
reCurse: Maybe it works better, maybe it doesn't
MSmits: so i should probably just try it
MSmits: cant hurt
reCurse: You need to threshold that with number of comparisons possible
reCurse: Whether you need to compute more possibilities than accurate eval
MSmits: more possibilities?
reCurse: Nodes
MSmits: oh ok
MSmits: what kind of last activation function would a seed predictor use?
reCurse: Linear
MSmits: ah ok
reCurse: Unless you don't care about precision at the extremes
reCurse: Then sigmoid, tanh, etc
reCurse: don't care = care
reCurse: Wait that was right
reCurse: I hate not having edits...
jacek: i would care less
reCurse: I usually type first and correct stuff later
jacek: ohai
MSmits: sigmoid/tanh has asymptotic behavior, so less precision at edges, its clear
jacek: good for you
MSmits: thanks for clearing some of this up reCurse
MSmits: going back to practicing with basics in my nnfs book :)
jacek: MSmits i only did eval the normal way, i still cant picture the seeds thing
MSmits: it's a bit complicated, lets say if i get it to work, i will write it up, if not, you can forget it :)
jacek: as first thing i would stick to the basics and if it works, then would try some more exotic approach
reCurse: Yeah the important is to get anything basic to work
reCurse: You can add the fancy stuff later and use that as a regression test
CamTheHelpDesk: anyone know if there's more than 5 homepage paths
jacek: if you can get the eval right, you then know your NN works
reCurse: I feel like we're having conversations in the playground sometimes
MSmits: to me the seeds thing isnt really more complicated, its just different. I used it before
MSmits: but harder to explain
jacek: if the seeds thing wouldnt work, you wouldnt know if its faulty NN or other bugs
MSmits: oh, because i already know other approaches work from you guys
MSmits: good point
MSmits: but i might start with a simple ttt or something
jacek: :+1:
fenrir: with NN: always start simple first, because even with huge bug, the NN may workaround a lot of issue and hide it
MSmits: btw dots and boxes has this too. Scoreless states where only the board matters and the net score from the rest of the game. That's where i got the idea from
fenrir: I had a bot in SC21 that was quite ok (~50 early legend) trained in selfplay, where both player had the exact same inputs ...
Gumarkamole: Does someone have issues with Kotlin too? My code passes all the test but after pending results I have 0%
fenrir: (ie player 2 was seeing the input of player 1)
MSmits: fenrir yeah a lot of times re curse/jace k and robo told of bugs they discovered very late
fenrir: but it didn't prevent it of 'working not so bad'
MSmits: when the nn was already working
reCurse: Oh boy yeah that sounds all too familiar
fenrir: :)
RoboStac: if you do try it on TTT first be very careful with tree search methods as they'll find endgame states so quickly the nn doesn't matter much
fenrir: I know I have to look for this bug (not the first time), but I have still bugged it ;)
MSmits: yeah that was my worry RoboStac, maybe i will use it as a policy nn only, 9 moves available = 9 outputs
MSmits: no search
MSmits: i think jacek did this
Darkboy: Has any one seen this code
reCurse: I still think CSB runner is the best
reCurse: I can hardly find an example where it's complicated enough yet very easy to figure out if it's right
MSmits: for some reason it feels very complicated to me reCurse
jacek: MSmits i did 1-ply search
RoboStac: I've still never got anything useful to work on csb
RoboStac: probably should try that again at some point
MSmits: oh jacek you just took the best out of 9 value lookups?
fenrir: a runner alone in CSB is quite easy to do and yes a good training example
jacek: MSmits yes
fenrir: I started with tha, I think pb4 also (but unsure, don't want to speak at his place)
MSmits: that would probably be easi
RoboStac: last time I tried my runner learnt to only turn one direction and just kept varying velocity to make the circles bigger / smaller
MSmits: +er
reCurse: Oh yeah that rings a bell
reCurse: lol
fenrir: if RL, the entropy regulrization is really really mandatory
MSmits: thats what the first drawings of my daughter looked like RoboStac :)
fenrir: otherwise very fast convergence to a minimum
fenrir: I mean local minimum (an a bad one)
reCurse: Maybe I'll write a CSB runner tutorial at some point... don't hold your breath though
reCurse: I think it's a great introduction to RL in general, CG regardless
MSmits: hmm, you're not the first to say that, but there is also bias since that is how you got into it
RoboStac: it was a long time ago though, probably should try it again
reCurse: Of course there's bias
reCurse: I've also seen what else is out there
jacek: nn has bias, its a good thing
fenrir: there is a game with a runner only in CG
fenrir: SL no ?
jacek: search race?
fenrir: yep
fenrir: I think that could be a good way to test a runner only
fenrir: here I did a DDPG
darkhorse64: SF2442 ?
MSmits: ye innory also said csb is better than oware to start with
reCurse: I'm pretty sure he meant csb runner
reCurse: Real csb is a whole other story
MSmits: ah ok
reCurse: Like not even close
MSmits: that makes sense
MSmits: so that other optim then
MSmits: search race
reCurse: Yes and no
MSmits: how so
reCurse: If you get it to run under CG somehow, sure
reCurse: Much easier to get it running with whatever you want without constriants
MSmits: oh you mean locall
reCurse: The viewer is much easier to recode than working under CG constraints
MSmits: ah right
MSmits: yeah, seems like a nice project
reCurse: Oware is simple enough, problem is it's very difficult to see if it's right
reCurse: Unless you spend time studying the game
MSmits: mmh i have my meta mcts
reCurse: CSB? There's a checkpoint.
reCurse: If it doesn't drive towards it, it's a complete failure
MSmits: yeah much easier to see
MSmits: well I have a ton of time between now and around september, I hope to get at least a good bot in 1 arena. Will not start any other project before i do. Think thats realistic?
KalamariKing: Thats a lot of time, I'd say so
MSmits: cool
MSmits: i know reCurse spent many months also, but I think I have more help/resources now, mainly people telling me what works and what doesnt
fenrir: in CSB, my first NN was a search without doing the gradient, I don't remember the exact name, but basically it's a perurbation of the NN coffeicent, accumulate a bunch of run, and modify in the righ direction
MSmits: no gradient?
MSmits: oh just adjust weights by a fraction?
MSmits: like 1/1000 of the weight value
MSmits: the nnfs book mentions that as a first order improvement over complete randomly generated weights
fenrir: it was an algorithm search when the function is not derivable
MSmits: nice if you got that to work
Wontonimo: i used TPS to do that with very good results.
fenrir: it kind of worked, not fast but it idid
MSmits: TPS?
Wontonimo: sorry TPE
Wontonimo: mybad
MSmits: TPE?
Wontonimo: aka Hyperopt
Wontonimo: http://hyperopt.github.io/hyperopt/
MSmits: ye google found it fast:
struct: Hi MSmits I also bought the book
struct: Takes me a while to understand some stuff though
ill-be-waiting-outside: hi
MSmits: I have trouble with the backpropagation part. Not because i dont understand the math, but because it's so dry. I guess there's no way to make that more interesting though. Other than that I had no problem so far.
MSmits: let me know if you have questions
MSmits: i didnt go beyond about page 200
MSmits: (yet)
struct: Im entering chapter 5 atm (page 111) error loss
MSmits: ahh ok
struct: Sometimes I just have to read multiple times to understand
struct: Because this is brand new stuff
MSmits: there are videos too
MSmits: they end around there
MSmits: maybe a bit further, i think loss is in there
MSmits: backpropagation isnt
struct: yeah I saw up to softmax
MSmits: ah yes, there's 1 more after
MSmits: softmax is 6 i think, there's a 7th
struct: ill watch it too
davilla: which fields in the description are visible when the class is a reverse?
davilla: Input? Output? Constraints?
davilla: Or is it just the 4 examples?
MrPapaya: none
MrPapaya: only tests
davilla: cool thanks
Wontonimo: anyone give any thoughts to which multi they are planning on attacking with NN first?
Wontonimo: it would be great if we all dogpile the same one
Wontonimo: oware-abapa was mentioned
VizGhar: So this is the site, where future machine learning experts will be raised?
ZarthaxX: and you will get smashed in future contests, prob yes
KalamariKing: oware sounds like a good nn challenge
Kramerius: i just joined this. 9th grade, first coding class. This is scary
KalamariKing: say, why are we all learning nns this morning
KalamariKing: Kramerius That's pretty epic. What exactly is scary?
struct: I dont expect to make one any time soon
Kramerius: no idea what half of this means
KelvinAndHubbles: actually yeah, theres only like 10 input nodes for oware, might make one and run it through a GANN
struct: I struggled quite a bit on early chapters
VizGhar: you mean chat Kramerius?
Kramerius: and yes the chat too
Kramerius: i think before i do anything, ill just spend some time looking at the documentation
KalamariKing: chats really advanced today I'll give you that
KalamariKing: I'm somewhat experienced in nns so I can at least contribute today
Wontonimo: NeuralNetworks
KalamariKing: Plural of NN, neural network
KalamariKing: Machine learning (think ai)
KalamariKing: Kramerius honestly experience in the field is better. The practices might be too hard if you're just starting out, so play a few clashes or smth
VizGhar: Kramerius ignore chat :) only open it if you have specific question... You will get lost here immediately
Kramerius: i think ive seen a video on that
MSmits: really dont worry about this Kramerius, it's not always like this. Just have some fun with puzzles. That's what i was doing on CG 3 yrs ago
ZarthaxX: so funny that recurs win with nn and now everyone is on that boat
MSmits: blissfully ignorant of everything to do with bots
ZarthaxX: following the winner :rofl:
ZarthaxX: won*
Wontonimo: didn't that happen with MCTS also
KalamariKing: I remember someone telling me that a nn would be irrational
MSmits: ZarthaxX nahh, this has been going on for over a year at least in my case
ZarthaxX: not for you smits
MSmits: did you not see me get wrecked in every leaderboard :P
VizGhar: ++ just play easy puzzles, read forums about solutions and maybe a bit of CoC
ZarthaxX: but everyone else lol
KalamariKing: So I wasted all of my time learning search algos and stuff, when I already knew nns
MSmits: ahh ok
Kramerius: On youtube, there was a guy who used NNS to beat Falppy bird or something
ZarthaxX: i saw your struggle
ZarthaxX: fighting with your handmade nns :P
struct: ZarthaxX I wanted to try it before, but never started it
MSmits: lol
ZarthaxX: struct but this was like a motivator right haha
struct: Even relu took me a while to understand
MSmits: Kramerius yeah flappy bird is one of the things you can do
Wontonimo: I used scratch.mit.edu to do RL for a self driving car
Wontonimo: well, it was a dog
RoboStac: I think it happens to some extent after most contests, just in this case it's getting discussion in chat as it's a bit more complicated / unusual (and something that the regular chatters haven't done before)
Wontonimo: not a car
MSmits: self driving dog?
MSmits: or a dog learnign to drive?
MSmits: a dog on wheels?
RoboStac: a dog learning to write a nn
KalamariKing: Or a car learning to dog
MSmits: oh ofc
MSmits: btw, scratch wow... talk about low tech :)
ZarthaxX: scratch rl?
ZarthaxX: how can you do that
MSmits: thats what i did with students before i could even code
ZarthaxX: its scratch, block programming
ZarthaxX: wtf?
jacek: KalamariKing you still need search though
KalamariKing: jacek yeah ig
maazshakeel: Hi guys
KalamariKing: Hello
ill-be-waiting-outside: hey
maazshakeel: How are you all?
Hasuzawa: just curious, can I use numpy of python in codingame, I figured the quickest solution to a problem is n by m linear equation with excess condition
jacek: they are up to something
MSmits: yes Hasuzawa i think numpy is an option
jacek: yes, there is numpy in here
jacek: and panda i think
Hasuzawa: cool, thx
MSmits: :panda_face:
Hasuzawa: now I dont have to use crammer rule
AntiSquid: ZarthaxX i think for this contest specifically it can work better, doubt it works for every contest
AntiSquid: for example: pacman
AntiSquid: at least it's far more difficult
Wontonimo: i just had to double check on the self driving dog. not rl, it was hand tuned. :( I didn't lie, I just remembered wrong
AntiSquid: here it was favoring optimizing eval and since that's what NN is good at of course there was such a massive lead with a well trained model
MSmits: I dont think you need to use NN for the entire thing. You can take a part of the game thats hard to evaluate and train a NN to do it
jacek: no, lets NN everything
AntiSquid: i guess, you can do "NN" for the RPS minigame part, or use the already available algorithms online for it#
MSmits: allright, I'm convinced
MSmits: let's nn everything!
KalamariKing: I agree
MSmits: oh btw, jacek, wanna see how far i got with connect4 so far. I recently turned on a 10 game paralllel version of the meta mcts, building on the old version
AntiSquid: VizGhar not really, this isn't an ML focused site, i am not an expert to say who would be best where, but i am sure the folks that train models all the time would have an advantage over everyone here
Wontonimo: i'm looking forward to training my NN with MCTS, then using that trained NN in my MCTS.
jacek: and how far did you go
MSmits: http://chat.codingame.com/pastebin/fae6edbe-0eb4-4ed5-b242-f24fceb2be1f
MSmits: for a while i did just move 1, it's doing all again now
jacek: :scream:
MSmits: thats why move 1 has 10 million games
MSmits: pretty sure it's p2, draw, p1, p1, p1 when it gets solved
jacek: but theres summer soon
KalamariKing: I was summoned
AntiSquid: is that with steal included MSmits ?
KalamariKing: Where was the ping wtf
MSmits: this is without steal, so this tells you to start with move 1 or get a forced loss as p1
MSmits: because move1 is pretty sure to end in draw. The search hits leafs that are solved in draw with the principal variation
AntiSquid: 51.7% so p2 is always at advantage with the steal ... so like in the popular 7x6 version ? :D the first move has the advantage
MSmits: no, the 51,7% goes above and below 50%
MSmits: it needs to completely solve lost branches too
MSmits: so that affects the percentages
MSmits: just in case a bad branch still leads to win (otherwise it is not a proof)
MSmits: i should probably stop backpropagating when there is a draw option. That would solve the fluctuation
MSmits: I kinda want to go all the way and solve this, because wiki says noone has yet
AntiSquid: ya definitely some branches can lead to a win, they do for 6x7 (i just know trust me0
MSmits: 8x8 is biggest solve, 9x7 would be next
AntiSquid: saw chat earlier, anyone who doesn't have the nnfs book ? :D
MSmits: not me, I don't not have it
AntiSquid: what are you reading then?
MSmits: nnfs
MSmits: because i have it
AntiSquid: @_@
jacek: neural network for file system?
MSmits: neural network from scratch (in python)
AntiSquid: McBitz jokes, where's Automaton2000
Automaton2000: but it will be worse
MSmits: cant me worse than mine Automaton2000
MSmits: be
Automaton2000: don't think he can do
jacek: or AutomatonNN?
KalamariKing: I think AutomatonNN is no longer with us, succeeded by Automaton2000
Automaton2000: maybe i should read the rules again
MSmits: Automaton2000 dont read the rules and use pruneArray
Automaton2000: if you go to the chat
AntiSquid: Automaton2000 is the original one, KalamariKing, well this is second iteration
KalamariKing: I'm in the chat rn what do you want Automaton2000
Automaton2000: we do have a lot of bugs in your code
Automaton2000: but when i try to make my code cleaner
KalamariKing: AntiSquid oh sick
KalamariKing: thanks
MSmits: wait, is he telling us he saw our code and it has bugs?
KalamariKing: Automaton2000 why would you try to clean buggy code
Automaton2000: what you need to put in some time
AntiSquid: use the search bar on the site to search for automaton, KalamariKing, check the puzzles out
KalamariKing: Automaton2000 if my code is as buggy as you said, then fine
Automaton2000: but i'm saying i like do actions like harvest, seeding
KalamariKing: AntiSquid I see the 'build your own Automaton2000' rn, its actually pretty cool
Automaton2000: makes me want to puke
MSmits: btw, this C# mutex thing is awesome. So easy. I think Astrobytes suggest this? Or someone else. Anyways thanks
KalamariKing: wdym mutex
MSmits: for parallelization, so you dont get many threads accessing and modifying the same variables
jacek: mutex eh
jacek: and virtual loss?
MSmits: i have 10 visits, not sure if thats optimal
MSmits: 10 vists for one that is being played through
MSmits: or if more than once 10x
jacek: per thread?
MSmits: yeah
MSmits: if multiple threads go through same node
KalamariKing: when did I reach bronze in two games
MSmits: then that node has x times 10 visits in UCT
KalamariKing: I've only played one bot game, csb
KalamariKing: just got a random acheivment
jacek: no spring challenge?
KalamariKing: ohhhhh
KalamariKing: that makes sense thanks
hypers: heh, I've got achievement for two golds and i'm not in the gold in spring challenge
hypers: just in one more contest as far as a remember
hypers: also happend minutes ago when rating was updated
hypers: I hope they wouldn't change their mind
hypers: as i'm already proud of myself :)))
KalamariKing: rating was updated a few mins ago?
KalamariKing: interesting
hypers: as far as I can tell by Clash of Code to
KalamariKing: Oh sick
hypers: i was not in the 500 this morning, now I've moved up as a result of today's games
hypers: but I expected update to happen tomorrow
KalamariKing: I thought your rating updates after every game
KalamariKing: It does for me
MSmits: the CP ranking updates at night
struct: coc ranks update after a game is complete
struct: But CP is daily yeah
hypers: I might have looked at CODINPOINTS
MSmits: thats cp
hypers: oh, thanks I see now
asrinivasan007: hello
asrinivasan007: orz
jacek: good afternoon
asrinivasan007: :wave:
Smelty: ^
KalamariKing: you said that literally 6 minutes before noon here
KalamariKing: could you not have waited
jacek: no :imp:
Smelty: lol
Smelty: hehe https://www.codingame.com/share-replay/559010104
jacek: oO
bora399: I need to improve my coding skills, any suggestions ? I can't solve medium puzzles
KalamariKing: bora399 The puzzle difficulty isn't consistent, I don't blame you
KalamariKing: Some mediums are easy, others are like expert+
bora399: oh okay thx for message, i relaxed :D
Wontonimo: some of the medium puzzles take me several hours, and many times I have to think on it casually for a day or two.
Smelty: yep
struct: This is why I dont do puzzles
Smelty: some hard puzzles r easier than some medium puzzles
jacek: never run with scissors
King_Coda: How do you hide chat using the url?
struct: ?disableChat
King_Coda: isn't it like ?hideChat
jacek: alt+F4
King_Coda: OH YEAH, thanksstruct
jacek: my method would also hide the chat
Wontonimo: i use your method often. it works well
Wontonimo: got2go :wave: l8r
Soumyadarshan: hey
jacek: good afternoon
Smelty: good morn
Wiilss: hi
Wiilss: :relieved:
Illusive: so can someone tell me why its called console.log, like bruh log is a tree
Wiilss: damnnn
Wiilss: facts
Illusive: if australias 12 hours ahead, why didnt they warn us about 911
551D: tree is a log but not every log is tree
Wiilss: twisty
AntiSquid: Illusive funny, what about "log a call" or "log in" ?
eulerscheZahl: hey, it's somewhat working \o/ https://eulerschezahl.github.io/AI-Battle/connect-4
struct: 404
eulerscheZahl: (╯°□°)╯︵ ┻━┻
eulerscheZahl: https://eulerschezahl.github.io/AI-Battle/index.html
eulerscheZahl: and then navigate on the left
OutLaw: hey everyone, I have two questions: is it possible to put the dark mode in all the codingame website, and where can I learn to code in C++? From the basics?
AntiSquid: sorry there's nothing at this address
jacek: weird, when i click the link its 404
jacek: but from home page its alright
eulerscheZahl: blazor single page application combined with github pages
struct: https://i.imgur.com/FQtZAlO.png
eulerscheZahl: there must be a way to get around this
struct: how many rollouts?
eulerscheZahl: i don't print it. also depends on your browser
eulerscheZahl: all done in WebAssembly
jacek: its not c#?
jacek: c# to js?
eulerscheZahl: yes, compiled to WASM
eulerscheZahl: no JS
Marchete: C# to js? amazing
jacek: no steal move?
jacek: now do onitama :imp:
eulerscheZahl: but that's a lot more work to do :scream:
jrke: euler here is mine connect 4-
jrke: https://github.com/Jrkej/Connect-4.git
jrke: ;)
eulerscheZahl: that's a lot of instructions compared to clicking a link and playing in the browser
eulerscheZahl: after fighting the 404
jrke: yeah but i made that for fun and school project
jacek: so c# bots on cg can be automatically converted?
eulerscheZahl: with very little effort
darkhorse64: I won it !
eulerscheZahl: struct too
ZarthaxX: where is the bitboard jrke?
ZarthaxX: no bitboards :C
struct: euler I would change the yellow pieces
struct: hard to see them imo
eulerscheZahl: just go and buy a better screen
ZarthaxX: toad i cant see your game
ZarthaxX: :(
struct: its not the screen
eulerscheZahl: did you try https://eulerschezahl.github.io/AI-Battle/index.html ?
AntiSquid: resizing width breaks the layout euler
struct: is just that night light is on
struct: to save my eyes
eulerscheZahl: or why can't you see it ZarthaxX?
jacek: he has screen off
eulerscheZahl: that's something I started yesterday and you guys are already complaining :(
ZarthaxX: yeah what jacek said, i didnt turn pc on
ZarthaxX: works like a charm toad <3
struct: im not complaining :/
eulerscheZahl: so, what else if not yellow?
eulerscheZahl: blue? :P
struct: well you can change the background
AntiSquid: poor web design needs to be thoroughly scrutinized, where's the contact us form to leave a review on the site?
ZarthaxX: did you do low lvl stuff toad?
ZarthaxX: :rofl: squido
eulerscheZahl: no, pretty much my tech.io bot still
eulerscheZahl: just with a larger board
ZarthaxX: oh okey
eulerscheZahl: and 400ms
ZarthaxX: as you said web asm
ZarthaxX: i didnt get it
ZarthaxX: i lost
eulerscheZahl: do you know what webasm is?
ZarthaxX: :D
ZarthaxX: i think i dont
ZarthaxX: i heard of it tho
eulerscheZahl: a way to run compiled programs in your browser
ZarthaxX: wait
ZarthaxX: you are telling me i can run a c++?
ZarthaxX: an exe
ZarthaxX: executable whatever
eulerscheZahl: and blazor provides a .net runtime as WASM
ZarthaxX: interesting
ZarthaxX: i wanted to try running bots on web
eulerscheZahl: so the connect4 bot is a .net DLL which is executed in .net runtime via webasm
ZarthaxX: not a good idea i guess tho
ZarthaxX: damn
UUolfo: Side note 2nd?!? dangggggg whose first and how do i tell?
ZarthaxX: interesting
eulerscheZahl: https://www.codingame.com/leaderboards/general/global
eulerscheZahl: royale reclaimed the lead
ZarthaxX: probably clashes e.e
eulerscheZahl: i dropped in the spring 2021 leaderboard
UUolfo: looks like bot programming
AntiSquid: 4 CP difference ZarthaxX
eulerscheZahl: 104
AntiSquid: 104 CP nvm
ZarthaxX: lol
AntiSquid: ya basically clash diff by 100 points
jrke: damn dbdr also in the race
eulerscheZahl: he will take over soon
eulerscheZahl: his points are from games that still gain more players and total points
jrke: yeah one those are d&b
ZarthaxX: true
jrke: and i am also a guru now :)
Wiilss: hi
jrke: just a rank above ZarthaxX
ZarthaxX: you took my place jrke
ZarthaxX: lol
ZarthaxX: right
ZarthaxX: i will fix it, someday
eulerscheZahl: was about to mention that
jrke: time to sleep... gn bye and hoping ZarthaxX will fix his rank
ZarthaxX: haha nah
ZarthaxX: not for now
ZarthaxX: gn!
AntiSquid: when you going to bed ZarthaxX ?
eulerscheZahl: in 10 hours
eulerscheZahl: it's 14:30 for him
eulerscheZahl: wait 14:53
AntiSquid: 1 AM ?
jacek: and when you goind to bed, euler
MSmits: in 1 hr
eulerscheZahl: in 1h
eulerscheZahl: (╯°□°)╯︵ ┻━┻
MSmits: and 7 minutes
AntiSquid: :D
MSmits: euler has perfect self control and self discipline
AntiSquid: it's usually 8 AM here when he goes to bed, makes it look weird
ZarthaxX: i hate this toad, always knows my time
AntiSquid: 8 PM *
eulerscheZahl: i still feel a little tired after the contest
ZarthaxX: who doesnt?
ZarthaxX: :D
ZarthaxX: they are getting more intense
eulerscheZahl: but long weekend ahead :ship:
ZarthaxX: true :O
jacek: long?
eulerscheZahl: 3 days weekends are the best
MSmits: whats with the boat
AntiSquid: you can search the time online ZarthaxX : https://duckduckgo.com/?q=time+arhentina
MSmits: you going on the water?
ZarthaxX: he is a toad
eulerscheZahl: like "land in sight"
MSmits: toadboat?
MSmits: oh ok
ZarthaxX: i know AntiSquid
eulerscheZahl: just that the land is a public holiday
ZarthaxX: but i think toad knows it without googling it
MSmits: right, totally get this analogy now
ZarthaxX: he says it too fast always lol
eulerscheZahl: i know ZarthaxX = me -5h
ZarthaxX: yeah that :P
ZarthaxX: i will remember the 5
ZarthaxX: wait, you go to sleep at nine???
ZarthaxX: or at 10
eulerscheZahl: 9 sounds about right
ZarthaxX: isnt that way too early?
ZarthaxX: why not 00 D:
MSmits: he gets up early
MSmits: feeds his chickens and stuff
eulerscheZahl: around 6am
ZarthaxX: hell
ZarthaxX: ahhh
AntiSquid: what about daylight savings ? are those always synchronized ?
ZarthaxX: makes sense
ZarthaxX: why tho?
MSmits: cuz they need food
jacek: he needs his dose of lucid dreaming
ZarthaxX: MSmits toad feeding chickens :rofl:
ZarthaxX: suddenly he lives in a farm
MSmits: i dunno why, it's just educated guessing
ZarthaxX: still no phone toad right?
ZarthaxX: no russian telegram
eulerscheZahl: i have landline
MSmits: me too, no phone
ZarthaxX: i meant mobile phone
MSmits: and landline
MSmits: yeah
ZarthaxX: isnt landline dying?
AntiSquid: my smartphone is my landline
MSmits: i dunno. It worked ok last time i tried
ZarthaxX: i meant like
ZarthaxX: passing away
eulerscheZahl: i also have a headset for the computer
ZarthaxX: lol
MSmits: me too, headset
eulerscheZahl: belongs to employer but i think i can risk it :imp:
ZarthaxX: :rofl:
ZarthaxX: computer also belongs to employer?
MSmits: i dont see the point of mobile phone
MSmits: isn't that for like, a social life
eulerscheZahl: using my personal screen and keyboard for work in return
eulerscheZahl: i have a computer on my own and one from the company
eulerscheZahl: i way i'm willing to use that windows in my free time
ZarthaxX: notebook from company i guess?
MSmits: same, 2 computers, one for myself and 1 for work
AntiSquid: i have a work laptop, they kept promising work phones for a while now (smart phones too)
MSmits: I should start playing soccer and then I'll live eulers life
eulerscheZahl: not a notebook. more like a 20kg monster
ZarthaxX: :rofl:
eulerscheZahl: 60cm deep
ZarthaxX: thought it was normal that they gave notebooks only
eulerscheZahl: and 40 high, 20 wide
ZarthaxX: you can train your NNs there :D
ZarthaxX: or mine
eulerscheZahl: it has 2 processors with 8 cores each
MSmits: whoa
ZarthaxX: wtf, what do you do at your job
eulerscheZahl: FEM machine
ZarthaxX: MSmits i think he doesnt feed chickens
AntiSquid: using windows in freetime? :thinking: got tired of linux ?
eulerscheZahl: i don't really need all that parallel power. it's just that this PC was spare when I needed one
ZarthaxX: okey :thinking:
MSmits: ah right, it was lyin' around
MSmits: makes total sense now
ZarthaxX: lol
AntiSquid: it's not about what you need it for, it's what the employer can afford
eulerscheZahl: definitely more expensive than my personal computer
AntiSquid: is it secured? encrypted?
AntiSquid: how much access do you have to it
eulerscheZahl: questions getting weird, not sure if that's the right topic for public chat
MSmits: why do you need to know that :P
AntiSquid: why not? curious
MSmits: this reminds me of conversations with mad knight :P
eulerscheZahl: let's do a teamviewer session
MSmits: come get on teamviewer and when i ask why, why not?
eulerscheZahl: :rofl:
MSmits: hehe
ZarthaxX: lol
AntiSquid: completely different topic
MSmits: my point exactly :P
AntiSquid: madao feels lonely likes to chat, different interest
eulerscheZahl: once another user here asked me to connect to his PC with teamviewer
AntiSquid: madao's alt?
MSmits: well that doesnt surprise me... you help people with technical problems
eulerscheZahl: no, i could help him with a C# problem
AntiSquid: so did you connect?
eulerscheZahl: but i wouldn't give any of you access to my PC. you are still somewhat strangers
eulerscheZahl: yes, i connected to his PC
eulerscheZahl: and i don't want anyone to touch my PC
AntiSquid: i didn't ask for access to your PC, weird assumption
MSmits: wouldnt do that either i think, but if anyone, it would be you
MSmits: nahhh it would be astrobytes
MSmits: :P
eulerscheZahl: just saying because that random guy on chat (not even a regular user) asked me
struct: at most screenshare
eulerscheZahl: if i were in his shoes, i wouldn't
AntiSquid: nobody here, i don't care, it's private @_@
BinaryIO: eulerscheZahl I have your location. locking in on you!
eulerscheZahl: because I wrote the location on my profile
AntiSquid: besides i have funny gif desktop can't have people see that
eulerscheZahl: you are in *concentrated look* the US
eulerscheZahl: i have a hypnotoad picture collection
Smelty: lol
AntiSquid: euler is a foreign spy, BinaryIO
MSmits: we're all foreign
Marchete: ALL GLORY TO THE HYPNOTOAD
Smelty: ^
MSmits: did you watch that 10 hr YT again Marchete
Marchete: ofc
MSmits: I know it only seems like 2 minutes, but watch the sun before and after
Marchete: twice in a row
eulerscheZahl: i hear that every day when writing extensions for the CAD software: "foreign application failed to initialize"
AntiSquid: oh god MSmits, just reminds me reading some ingame chat, often seeing "oh all these foreigners not speaking my language, get out of my country" lmao
MSmits: which game was this AntiSquid?
AntiSquid: happened in multiple ones
MSmits: yeah, I remember chat like that in dota2
AntiSquid: "verdamte auslaender"
eulerscheZahl: verdammte
MSmits: they always complained about "the russians"
AntiSquid: ya that
eulerscheZahl: verdamt = something that turns into a woman
Smelty: :eyes:
reCurse: I am now fluent with insults in spanish, portuguese and russian thanks to dota 2
MSmits: lol reCurse
struct: lol
Smelty: lol
eulerscheZahl: i know some Hindi curses thanks to kitboga
struct: brazilians play it?
MSmits: i hadnt watched kitboga in months and i coincidentally watched 1 last night
MSmits: i think you need, to every once in a while
AntiSquid: what MMR reCurse ?
reCurse: Between 3.8k and 4.2k back when pros were around... 7 or 8k I think?
reCurse: That was years ago
MSmits: I never got very far with that, you had to play hundreds of games to even play rated games
reCurse: Leagues didn't even exist
MSmits: i stopped when i was allowed to
AntiSquid: i only played it back when it was a warcraft 3 map
MSmits: oh thats a long time ago
MSmits: like 15 yrs ago?
MSmits: more even
AntiSquid: idk, it was getting popular on ggarena or something like that
AntiSquid: gggaming !
MSmits: only game i played seriously was company of heroes
MSmits: online i mean
MSmits: lots of strategy, people around my rating had 3x as many clicks/min
MSmits: slow but smart :P
reCurse: APM got perverted when people started putting too much focus on it
AntiSquid: high APM = pro
MSmits: depends on the game though. Some games are specifically designed to require it, like starcraft 2
reCurse: Same result
AntiSquid: 1212121 right click spam
reCurse: Hardly measures anything except maybe sweatiness
reCurse: Of course you need some good reflexes to be good but
reCurse: Past a certain point...
reCurse: Just realized Nerchio might be reading this, I should shut up in case I say something very stupid
MSmits: well, with that logic most of us should keep quiet because you're here :P
reCurse: Um...
reCurse: No?
AntiSquid: nerchio teach us the right hotkey bindings
MSmits: I mean because i say a lot of stuff you know much more about
MSmits: sometimes saying stupid stuff is an opportunity to be corrected :)
reCurse: Fine, let's just say I might be speaking with a bit too much confidence when I could get massively corrected by someone who (was?) an actual pro
MSmits: yeah, but you can take the embarassment, it's fine
AntiSquid: you can check if he is in chat
AntiSquid: protoss are easy win because shields OP
MSmits: he's not, but we log chat these days
MSmits: it's actually way more risky for me than for you reCurse
MSmits: i am a teacher, i often think sht, some student may read this
AntiSquid: see? he didn't respond
reCurse: Alright so excluding actual pros
MSmits: if i am careful half the jokes i do i probably shouldnt :P
reCurse: APM is most likely perverted by unnecessary sweatiness in those levels below
reCurse: IMO
MSmits: do you mean perverted as in inflated?
reCurse: Yeah because it's often wrongly associated with a measure of skill
Dartisan: Hi, for those who interest, I put nine men's morris (Multiplayer) back in the contribution. ;-) Any feedback is welcome.
MSmits: ahh ok
reCurse: So people get it all wrong
reCurse: I need more APM to be good
reCurse: ...no
MSmits: so people are just randomly clicking stuff in downtime
AntiSquid: it's the culture, people measure APM and if it's not super high for no reason then they call you a noob
reCurse: Exactly
reCurse: That's royally stupid
reCurse: Sure you need to poll your buildings etc to make sure you build stuff on time and all
reCurse: But that's why I say past a certain point the metric is useless
SubwayMan: Hi, is there any way to adjust the codingame editor settings? I'm really not a fan of the autocomplete
UUolfo: I think it used to be that it was a measure of how percisely you could do x commands in y time, but it turned into just rando click
MSmits: games become a lot less interesting to me with these features. Like the last-hitting and deny-ing thing in dota2
MSmits: they could have coded a feature where this was automatic
MSmits: and you could focus on thinking
struct: denying is a joke
AntiSquid: like Heroes of the Storm ?
struct: dota gives me raic flashbacks
reCurse: That's massively underestimating all the strategy that goes with it MSmits
MSmits: i like that you dont have to do it in league of legends and hots
jacek: is it alive still?
reCurse: And it's not nearly as much about reflexes
AntiSquid: "pick heal" "omg noob why not pick heal"
MSmits: reCurse how so
reCurse: There's a lot of mental games that go with it, which creeps to focus, zoning, etc
struct: its more about timing decision
MSmits: i am ok with decisions like those
MSmits: it's the ridiculously precise timing required
AntiSquid: that's why you play 24/7
reCurse: There is a whole balance around attack animation and damage for last hitting and denying in dota
MSmits: yeah i know
struct: I tried to play dota but the learning curve was to big, compared to league
MSmits: i preferred heroes of the storm
MSmits: it seems a lot more random/fun
AntiSquid: extremely random
MSmits: yeah
MSmits: for casual game it's better i think
AntiSquid: i remember the fun times: ping once for help reply: "STFU YOU TOXIC SHIT"
MSmits: toxic?
MSmits: lol
reCurse: Ah yes HOTS, where you get called out for being toxic whenever you dare suggest a strategy to the team
AntiSquid: they loved the word
reCurse: How fast I quit that game
MSmits: ahh I dont remember, maybe i didnt play during the toxic era
struct: tbf I dont know any moba where there arent toxic ppl
reCurse: Sure, it's just the complete opposite in HOTS
AntiSquid: what league were you in MSmits ? @_@
MSmits: dont remember honestly, i just did daily quest stuff for a few weeks
King_Coda: That scroll bug is annoying
Murat_Eroglu: hey reCurse, do you have a github?
reCurse: In usual moba you get toxic people on you when you don't do things right
reCurse: In hots you get toxic people on you when you want to do things right
reCurse: And get called toxic for it
struct: ah lol
MSmits: I see
AntiSquid: "omg that noob picked x skill" "gg, let's abandon"
reCurse: Murat_Eroglu: Yes but not using it for anything
agysejt: Hi recurse! what was your method?
reCurse: ?
MSmits: pretty sure he refers to contest
MSmits: there's a post mortem agysejt
AntiSquid: blizz games have a weird community in general, i lost count of the times people added me in hearthstone just because they lost a game and felt angry
agysejt: BSearch or any method?
MSmits: he used a NN with duct i believe, but read PM
Murat_Eroglu: reCurse, where can I find your post mortems?
MSmits: many interesting PM's
jacek: https://www.codingame.com/forum/t/spring-challenge-2021-feedbacks-strategies/190849/67
agysejt: is any PM anywhere?
MSmits: Murat_Eroglu they are organised in threads per contest
jacek: (not rick roll)
agysejt: ohh thx
Murat_Eroglu: thanks
eulerscheZahl: https://www.codingame.com/forum/u/recurse/summary check the top replies as you said "post mortems" in prural
agysejt: anyway grat reCurse !
MSmits: oh nice one eulerscheZahl
Murat_Eroglu: thank you very mucj
Murat_Eroglu: much*
reCurse: Yeah check jacek's link that's where I wrote my postmortem
MSmits: eulers link is better
agysejt: thx , i see already
kovi: 129 hearths...
UUolfo: is there a way to see live code variables during operation instead of displaying it as a print?
MSmits: pure love kovi
eulerscheZahl: that's probably the most liked post on the entire website
eulerscheZahl: the petition to bring back code4life had 115 last time I checked
reCurse: :blush:
MSmits: lol
agysejt: what is the best c++ framework for AlphaZero (github example) ?
reCurse: I don't know of any
reCurse: It's usually all in python
MSmits: wait, now i am confused. Code 4 life is not on here anymore, but someone asked a code 4 life question during contest
MSmits: was this person trolling?
jacek: /me
MSmits: was it you again jacek?
kovi: yeah, shame there is no c++
kovi: but it probably wont be cg compatbile anyway
Dartisan: No Code4Lie is on, I am currently on it :-)
1415495: well 90% of the fun is to do it onself (IMHO)
kovi: ^
eulerscheZahl: code4life is here. but there was a delay of more than 1 year between contest and multiplayer
MSmits: ohh ok
jacek: during coitus also?
agysejt: yes i see many python examples
jacek: checking out leela zero code would be the closest
eulerscheZahl: jacek asking for moderation?
jacek: :no_mouth:
Marchete: alphazero.jl could be a good starting point too
Marchete: in julia
Marchete: https://jonathan-laurent.github.io/AlphaZero.jl/stable/
MSmits: thats the only azero i got to work locally actually
MSmits: the rest required versions of stuff i could not get working or a linux system
agysejt: ohh Julia another language?
eulerscheZahl: yes
MSmits: then i started training and it said wait 20 hours
MSmits: so...
reCurse: The alphazero implementations on github are very unsatisfying
reCurse: Those I found anyway
AntiSquid: you need to specify exact versions for each python library MSmits
MSmits: I know AntiSquid
MSmits: I dont have patience for that sort of stuff :(
MSmits: I used anaconda too, supposed to make it easier
eulerscheZahl: isn't that why we have virtualenv?
MSmits: still stuff goes wrong
MSmits: yeah i did that
MSmits: only got the julia one to work
Marchete: python code is usually hard to run
Marchete: incompatible versions
Marchete: esp. old tensorflow examples
reCurse: Especially if it depends on tensorflow, good luck
reCurse: lol
Marchete: :D
MSmits: i guess when you're a pro coder you're used to dealing with all this compatibility crap, but I dont really have the patience :)
reCurse: Tensorflow versioning is a clusterfuck
jacek: why do they deprecate stuff
Marchete: I totally agree
eulerscheZahl: i hated that tensorflow task on CG
AntiSquid: well if you run on windows 10 you sometimes need to figure out how to manually install specific libs yourself and in which folder to add ... lots of time spent on config, reason to use linux instead
eulerscheZahl: solved the puzzle by digging through older versions of the docs in git history and just copy-pasting that
agysejt: super :-D
MSmits: well you used what you're good at
MSmits: finding stuff that wasnt intended to be found :P
jacek: huh
kovi: i was lazy and used workspaces on udacity or on colab and keras wherever possible
agysejt: is something caffe c++ CNN library
agysejt: udacity or on colab?
AntiSquid: still have aws credits i need to use up @_@
kovi: udacity trainings...colleges who did local spent 30%+ time on environment/setup/compatibility issues
agysejt: i see, thx, something school
kovi: but no c++ nn
reCurse: Oh man, reminds me
reCurse: First day with my brand new rtx 3080
jacek: :unamused:
reCurse: Pytorch dies
reCurse: No support whatsoever for it
AntiSquid: where do you live?
Wontonimo: nice machine
reCurse: Had to hack around in source and building it from scratch to get it to work with latest cuda
reCurse: Fun day...
agysejt: when will decrease rtx price? :-D
AntiSquid: when the assembly line runs at full capacity again
reCurse: We can start by coming up with a plan to kill all cryptos
Wontonimo: and demand dries up
reCurse: I'm willing to dedicate time for it
jacek: all hope in musk
agysejt: Musk :-D
kovi: and altcoins drop a bit more
AntiSquid: crypto's here to stay
AntiSquid: very long term
jacek: i heard eth's mining will be over soon?
Wontonimo: we just need one quantum computing breakthrough and then mine all remaining coin in 0.012 sec
Wontonimo: then it's over
jacek: breakthrough :thinking:
MSmits: heh, by definition, will that not just cause more mining?
MSmits: it will just devaluate the coin
MSmits: oh you mean literally mine everything?
MSmits: not sure what the algo's are, but isn't it like prime numbers where you can go on forever?
jacek: http://www.righto.com/2014/09/mining-bitcoin-with-pencil-and-paper.html
jacek: btc supply is limited
agysejt: it is green method :-D
reCurse: I'm hoping regulation is how it's gonna get killed
reCurse: China had the right idea :P
MSmits: probably is less green than using a computer agysejt :P
kovi: gree: are you sure? considering the power input of people?
MSmits: ^
AntiSquid: CCP isn't the ideal gov
reCurse: That's not what I said
agysejt: too many paper and pencil? :-D
MSmits: they can still have good ideas
MSmits: even if they're genocidal
AntiSquid: and you'll probably get state run crypto at some point
MSmits: on the side
jacek: like china's chia
reCurse: Chia is a joke I think
jacek: which uses storage as PoW :thinking:
reCurse: Wouldn't be surprised if it's hard drive suppliers trying to create a demand
agysejt: anyway bitcoin or similar things totally are negative products
MSmits: agreed
MSmits: unless there is some entertainment purpose to them
jacek: at least it should be based on folding@home PoW or similar
jacek: some useful PoW
MSmits: pow?
jacek: proof of work
eulerscheZahl: potw
MSmits: ah ok, i only knew prisoner of war
reCurse: It's just an unregulated instrument of speculation
reCurse: The moment governments start responding to it I assume is when it starts dying
MSmits: people just don't understand you can't create value out of nothing. Every piece of profit created with bitcoin is someone else's loss
taha-lyousfi: wassup guys ! I am recording a video do you wanna say something to viewers ? ?
actualCoderTrevor: Hello World!
LeviathanProgramming: Interesting. I did not know about the bitcoin thing, MSmits
Marchete: from the sun
Marchete: you can create from the sun
LeviathanProgramming: light is created from the sun
LeviathanProgramming: as well as heat and radiation
taha-lyousfi: wassup guys ! I am recording a video do you wanna say something to viewers ? ?
MSmits: you mean you can mine coins on solar power Marchete?
jacek: MSmits tell that to feds printing moneyz
taha-lyousfi: hey
taha-lyousfi: wassup guys ! I am recording a video do you wanna say something to viewers ? ?
MSmits: heh yeah jacek
jacek: meow!
MSmits: woof
1415495: money creation is necessary, it's only when in excess that is dangerous
MSmits: thats the unregulated part
MSmits: (with bitcoin)
1415495: bitcoin doesn't create money in itself
MSmits: how do you mean
MSmits: it's used as money
1415495: only states (directly or delegated) can
1415495: that's not the same thing
MSmits: it's not the same no
1415495: the US/EU does create money from thin air (directly or by allowing banks)
1415495: bitcoin creation create no extra money in itself
MSmits: when it's mined, more of it is in circulation no?
1415495: people gives a value to bitcoin but that's all
MSmits: ahh of course but in a sense thats true for "real" money as well
1415495: they are no 'new' euro/dollar in existance whena bitcoin is created
MSmits: it's just backed up better
MSmits: but countries have suffered hyperinflation
1415495: not exactly when, a new dollar is created, it exists (you had N dollars, now N+1) but when a bitcoin is created N dollar dosn't changes
MSmits: like in world wars
BlackDereker: clash of code should have a mode for the most optimal code
BlackDereker: that would be neat
Bashar: i got a question
Bashar: is it possible to find the original string from list of subsequences made from that original string?
MSmits: eh, thats a weird question
MSmits: are the sub(strings?) ordered in some way
MSmits: i mean does every character exist no more than 1 time in this list?
MSmits: in the string i mean
MSmits: that way youcan start with a character and append or prepend characters, by going through the list
Bashar: all original string's characters are distinct. each string in the list is a Subsequence of length 3 of the original string
MSmits: ahh ok
struct: this sounds like a clash
MSmits: take the first item in the list as a base
MSmits: go to the next item, see if any character is the same, add stuff to the end of this character if it isnt already there
MSmits: keep going through the list, adding stuff and prepending stuff, till your string doesnt change anymore
MSmits: that should give you an idea on how to solve it I guess
MSmits: it's not an easy clash tbh.
MSmits: if it is one
Bashar: its not a clash
Bashar: its a question made by someone ik locally
struct: tesla roadster 0-100km/h in 1.1 sec o.o
Bashar: i will try thanks man
MSmits: np
struct: Seems like a bit too fast imo
LeviathanProgramming: wow, that's crazy speed
MSmits: crazy acceleration
jacek: 3*vel?
MSmits: speed is meh
struct: max speed is over 400kmh MSmits
MSmits: oh ok thats better
MSmits: but can it fly?
MSmits: other things that go that speed can fly
struct: with spaceX package
MSmits: so...
AAlbertini: Hi everyone, I just joined Codeingame & I'm trying the Temperatures exercise in Kotlin but it doesn't matter whatever I try my submit doesn't pass any test even for the empty temperatures array just return 0 nothing works but when I manually try the input from the validators it works. Is there an issue with kotlin ? Thanks for your help !
jacek: hmm
jacek: have you tried other language?
MSmits: doubt there is an issue with kotlin, but i cant help you with that language
jacek: someone complained about kotlin today
MSmits: oh
AAlbertini: I passed it in C# & C++
MSmits: hmm
MSmits: weird
MSmits: dont know any kotlin masters here
jacek: hmm compiling takes unsually long in kotlin
jacek: AAlbertini what do you mean by return 0?
LeviathanProgramming: javascript and python are really fast
LeviathanProgramming: typescript is super slow
AAlbertini: @jacek Well I created a computeClosestToZero returning an int and did a println in the main to display the output and even if I let it empty and just print 0 the "Displays 0 if no temperature" validator doesn't work either
reCurse: There's issues with Kotlin CG is working to fix it, last I heard
jacek: when i hardcode print it works for me anyway
jacek: there is only println, no return
reCurse: Something with compilation speed so I assume the more code the worst
AAlbertini: Oh ok thanks because it's for a test interview with Codingame (android dev) and I tried from the tutorial window with this exercise and retried on the original codeingame to see if there is a difference but failing in any case
reCurse: I have no idea about how things are on the "for work" side
jacek: they recently switched compiler for kotlin (in other words - kompiler) which is slower, but i dont know any more details
AAlbertini: Well it's just a simplified IDE with questions and exercises
AAlbertini: Yeah noticed kotlin is really slow. Thanks for the help anyway !
Bashar: @MSmits based on what should i prepend stuff?
MSmits: ok say you're looking at one of the substrings
MSmits: the third character
Bashar: can i dm you the code?
Bashar: my code
MSmits: and its the same as the first character of your string
MSmits: that means you can prepend the 1st and 2nd character of your substring to the string you're building
MSmits: i suck at reading other people's code, especially if it is a different language
MSmits: you're basically just matching substrings to the string you're building. the part you already have, you ignore. The stuff you dont have, you prepend and/or append
MSmits: and you keep doing that until you did a full loop through your list of substrings without prepending or appending
MSmits: then everything in there is in the full string
MSmits: i would just loop over a substring like this:
Bashar: http://chat.codingame.com/pastebin/1ed5ddf9-11a1-4fdb-80d5-3b12e5e5bb93
Bashar: this is what i wrote
MSmits: yeah thats not much yet, but hear me out, maybe this helps
MSmits: maybe we should go to pm
Bashar: okay
jacek: hello
Smelty: meow
Wontonimo: hey Bashar, just joining the conversation. Are you trying to add all unique numbers in a list?
Smelty: o.o
Wontonimo: O.O
jacek: ._.
Wontonimo: ô.o
Smelty: O.o
Wontonimo: taco
jacek: tacocat is a palindrome
Wontonimo: palindrome isn't unfortunately
Smelty: Aibohphobia is the fear of palindromes
Smelty: coincidentally, Aibohphobia is also a palindrome
jacek: and Hippopotomonstrosesquippedaliophobia is fear of long words
Smelty: yes
Wontonimo: i was just going to say
Wontonimo: had to google the spelling, but u b faster
Wontonimo: "smiles" is a longer word though
Smelty: Papaphobia is fear of the pope
Wontonimo: cuz there is a mile between every s
Wontonimo: lol
Smelty: lol
Wontonimo: *pulls up dad pants*
Smelty: Arachibutyrophobia – Fear of Peanut Butter Sticking to the Roof of One’s Mouth
Smelty: ....interesting...
Wontonimo: you know, cuz dads have their pants pull up half way to their armpits
Smelty: yes
Wontonimo: whatever
Smelty: *gives wonton a box full of tacos*
**Smelty says it bought it at the local store
Wontonimo: :thumbsup:
Smelty: ccllash
Wontonimo: wait! these tase suspiciously like the same ones from the cooler
**Smelty quakes in fear
**Smelty says it bought it from the tacoshack across the road
Wontonimo: what are you up to jacek ?
Riku5720: hello
jacek: ?
Wontonimo: hello
Smelty: h
Smelty: e l l o
Wontonimo: w
Smelty: ;(
Smelty: just lost a clash
Darkboy: H
Darkboy: e
Darkboy: l
Darkboy: l
Darkboy: o
Smelty:
Smelty: http://chat.codingame.com/pastebin/9f5ca3b1-5325-4ba8-9fa2-e7aa8bb74abd
Smelty:
Wontonimo: so long Smelty, that's for all the tacos
Astrobytes: I take a day or 2 away and this is what happens to the chat?
Smelty: :0]
Smelty: h-hi a-astro
Astrobytes: Fix your crap messages Smelty
Wontonimo: help us please
Smelty: s-sorry
Wontonimo: we are in a loop.
Wontonimo: what's a good topic Astrobyte
Astrobytes: Chat's on acid?
Smelty: meow
jacek: happy Caturday's eve
Smelty: oop meant yes
Astrobytes: likewise jacek
Astrobytes: Good topic - don't ask me, I'm leaving again in a minute
Wontonimo: i've started calling Thursday "Pre-Friday" so I don't feel bad about Pre-Friday drinking
Astrobytes: Perhaps something about bitboards, NNs or psychedelic drugs
Wontonimo: so, i'm not a fan of bitboards
Astrobytes: That's how it is in Scotland Wontonimo
Smelty: s
Astrobytes: It's the night you ramp up your intake in prep for Friday
Wontonimo: now, before you go and jump on that comment and praise bitboards let me be clear: They are super useful and to their thing
Wontonimo: I just don't look forward to coding it.
Astrobytes: They can be annoying though
struct: if it wasnt for bitboards I wouldnt have coded on last contest
struct: I cant stand hex grids
Astrobytes: Wontonimo: https://gekomad.github.io/Cinnamon/BitboardCalculator/
Wontonimo: oh, nice
Astrobytes: I always recommend your own printing method though, with tests for everything. Write once, reuse. Or modify.
jacek: you dont have bitboard in uttt, Wontonimo?
Wontonimo: if it got to 40k sims on move 2 without bitboard I'd be some kind of god. No, had to do bitboard. Absolutely
Wontonimo: it's super useful and a must for optimization. Just feels like work
Astrobytes: It's good to put in work sometimes. Reward yourself with extra tacos.
jacek: after a while writing board in classical way will feel more effort than bitboard
Wontonimo: hopefully. Than soon someone from work will come to my desk and ask me what the heck I just submitted to the main code base
Wontonimo: *then
Astrobytes: Anyway, I'm out. See ya tomorrow or Monday-Tuesday
struct: cya
Astrobytes: take care all
Wontonimo: take care
jacek: mods are gone, lets spam :imp:
Wontonimo: bwah hahaha!!
Wontonimo: the chat is ours
jacek: ohai
Smelty: oooh my
Smelty: hipopotamus'es
Voudrais: Też jestem z Gdańska
jacek: :+1:
Smelty: *e*
Riku5720: first W
Riku5720: wait nvm 2nd
Smelty: lol gj
PatrickMcGinnisII: Dice Duel runs slow as hell on my pc
PatrickMcGinnisII: sd or hs doesn't matter
PatrickMcGinnisII: or hd
Smelty: F
Smelty: it does for me too
PatrickMcGinnisII: I'm runnign 3 monitors, i duuno if the web 'helper' engine isn't set right or just the browser sux
PatrickMcGinnisII: seems to be running better today than yesterday
PatrickMcGinnisII: hmm, still slow tho
PatrickMcGinnisII: gave browser higher priority, still not great
PatrickMcGinnisII: any1 have any other suggestions, let me know
therealbeef: is that a CG game?
Smelty: yep
Smelty: https://www.codingame.com/multiplayer/bot-programming/dice-duel
Smelty: tbh babylon tower is the easiest legend achievement
therealbeef: wow, just the front page of that game, with the example game is making my cpu go to 30% cpu utilisation, and staying there even though the animation stopped
therealbeef: i'll play that game in winter ;-)
AntiSquid: 3D replay ! you can rotate it
AntiSquid: the table is basically a UFO
AntiSquid: https://imgur.com/a/B0dV3Ve
Smelty: yes
PatrickMcGinnisII: ends on no valid moves? should say boss won https://www.codingame.com/replay/559100585
Smelty: hmm
PatrickMcGinnisII: oh well, off for dinner
PatrickMcGinnisII: glhf
Smelty: yes
jacek: dinner?
AntiSquid: jacek how did you came to 30 features to use and how did you select those 30 is far more important than the number itself, you can have more than 30, not meaning to be abrupt, but the number itself says nothing .
jacek: trial and error and experimentation. the number could be higher if there was much more benefit to have more meaningful features. but i needed to keep it low due to CG size limit
jacek: 81 * 30 thats already 2430 inputs. if i have 32 hidden units, that gives 77.7k weights and each weight is 1 character, so that gives me around 22k for the rest of code
jacek: which of those features are important comes from experiments. unfortunately at this points new experiment would take hours to conduct to see if it yields better results than my best net
AntiSquid: cool, but there's a whole "area of research" on feature selection and analysis of their usefulness
jacek: it was more handcrafted than automated :c
AntiSquid: ah just wanted to say it's weird those guys are looking for 30, but they don't know which 30 they are looking for
AntiSquid: whichever works best for you is good, but there are methods to go about it, you might not even need 30, i don't know
jacek: let them roam :imp:
jacek: for all we know i could be trolling
Smelty: ah mind gams
Smelty: *games
Lexer31: meow how are you
Smelty: meow meow (good)
Smelty: also....31 followers wow
jacek: huh
Riku5720: heya
ZarthaxX: gz Smelty :)
asrinivasan007: :nerd:
Hydrazer: (╯°□°)╯︵ ┻━┻
Smelty: lol
zonbita: meo` moe' meoemeoeoemeeo
Smelty: yes finally broke the 2000 coc barrier
Smelty: hello
JohnCM: haha feels good using python again for shortest mode
JohnCM: i've been using C++ past few days just to challenge myself, and it's been rough
JohnCM: C++ still ok for fastest and reverse, just not shortest
Mailaanhem: I want to code better anh learn English
Guilherme300: I didn't know how to solve ANEO Sponsored Puzzle
Guilherme300: I don't*
KelvinAndHubbles: since therers a 36% success rate on that puzzle, I'd bet a ton of others don't know how to aswell
NotSureWhyThisWorks: also a heads up with that puzzle even though my solution passed all test cases it got 80% on submission. I found out that it was a rounding error and now it gets 100%