eulerscheZahl: objectively seen i should have never been there
eulerscheZahl: i don't care about the overall leaderboard
eulerscheZahl: only the contest leaderboard matters for me
eulerscheZahl: royale played a multiplayer, gained 800 points and now he's first. he could have done that earlier, not a matter of skill, just time invested
eulerscheZahl: no, others can improve too
eulerscheZahl: and the top is close https://www.codingame.com/leaderboards/general/global
mark7: What does dynamic sorting want to say. I can't understand the problem statement even.
mark7: if age is same then sort them by name, if name is same sort them by gender and if gender is same sort them by surname? Am i getting it correctly?
eulerscheZahl: yes. depending on the search query of course
mark7: Okay thanks :-)
mark7: Yay weekly problem solved :-)
MSmits: hey eulerscheZahl, did you see what i said about onitama last night?
MSmits: you'll like this
struct: What did you discover?
MSmits: ok so you know how I've been running thousands of CG bench games to fit constants?
MSmits: so I was wondering why I have a winrate of 80% in cg bench and only 60% or so on cg stats
MSmits: less vs karlis o and tric trac
MSmits: turns out, my json file was configured to only ever run seed 0
MSmits: so i got nr 1 overtraining on one seed :P
MSmits: the same 5 card always
Illedan: I felt that
MSmits: already getting better now, after having started with full random
MSmits: for example I did not have knight-count in my eval because it did nothing for seed 0, but at full random it seems to help a little
eulerscheZahl: that's a great way to use the CG servers :P
MSmits: lol... well it wasn't completely for nothing. Just really silly :P
eulerscheZahl: and still you somehow managed to get first
MSmits: yeah, I suppose different cards makes a differents but it's not everything
eulerscheZahl: on a coinflip with the other two at the top
MSmits: the other two?
eulerscheZahl: with a similar score to yours on the leaderboard
MSmits: oh, right
eulerscheZahl: i'm glad to see some activity for that game now. was dead before
MSmits: it's one of the best games on CG I think
MSmits: people just dont know it
MSmits: i enjoy it a lot
MSmits: has great depth
MSmits: randomness prevents hard coding
MSmits: and looks good
MadKnight: oh the contest starts in 6 days
struct: Will prizes be announced today?
eulerscheZahl: we don't know
MadKnight: what's the prize, Automaton2000 ?
Automaton2000: am i the only one
MSmits: you're not a prize
MadKnight: Automaton2000 says Automaton2000 is the prize
eulerscheZahl: win an Automaton2000
Automaton2000: so now i can do the same
Automaton2000: what are you doing atm marchete?
eulerscheZahl: where is Marchete?
MadKnight: will magus ship this prize ?
MSmits: no idea, Automaton2000 misses him
Automaton2000: wish there was a problem
MSmits: you wish there was a problem with what Automaton2000?
Automaton2000: if i can do it in c#
struct: Have you coded an Yinsh bot yet Automaton2000?
Automaton2000: how to do any of you solved the puzzle of the week?
LelouchVC2: where's that?
struct: puzzle of the week?
Illedan: (╯°□°）╯︵ ┻━┻
MSmits: why are you flipping out Illedan
Illedan: I can't figure out why me damn goalkeeper is hugging that ball
MSmits: it doesnt know whcih way to turn
MSmits: happens in csb
MSmits: depth issue
MSmits: turning in csb takes 10 turns for a full 180
Illedan: no turning here?
Illedan: Charge by angle..
MSmits: this is like meanmax?
Amirabbas: it would be great if there was a rematch button in the private clashes....i send them this request ...I hope they do that
MSmits: it;s awfully close to the goal
MSmits: floating point issue?
wlesavo: may be if you try to go negative coors it would be easier to grab
wlesavo: it is possible to print out coors beyond the board, dont know if it helps though
Illedan: Yeah, I know
wlesavo: anyway seems to be a rare case
eulerscheZahl: Illedan when you move towards it, you will push it in your own goal?
eulerscheZahl: or at the last the sim thinks it
RoboStac: I think you might just not be able to pick that up with the 8 angles - it's partly over the goal line so you'd have to hit it pretty exactly to cover it before rebounding away
MSmits: smart RoboStac
MSmits: that's it ofc
pb4: What 8 angles ?
RoboStac: he's only trying 8 movement angles
wlesavo: he simulates at 8 angles
pb4: oh right
MSmits: Illedan is doing what you told me to do a year ago pb4
pb4: I don't remember what I said :D
MSmits: you told me to use smitsimax on FB
pb4: oh right
MSmits: never got around to that
pb4: Probably a lot of room to improve on FB
MSmits: yeah, it's a messy game
MSmits: pretty hard, complicated sim too
pb4: my thing is still a GA, and it's kind of been shown on other games that there are much better approaches
MSmits: might not be smitsimax though, i had my doubts.
MSmits: but Illedan is doing ok so far
pb4: how ok ?
MSmits: top gold I think?
MSmits: not sure if he beat the boss already
pb4: @Illedan, for your problem, one possibility would be to add a special 9th action
pb4: 9th action would be "pick closest snaffle up" and would have some kind of logic/geometry to go as close as possible to the snaffle while avoiding the wall
MSmits: yeah thats close to what I suggested about scoring
MSmits: have an extra action with heuristics for a well aimed shot
MSmits: pb4 will you be in next contest as well?
MSmits: I know you are time restricted
pb4: Probably not, I've largely exhausted my wife's patience with ooc ^^
MSmits: thought so, good to see you in it though
MSmits: and grats on the win, not sure if i said that before
pb4: At first I justed wanted to toy with the tracker.... had no intention to go further -_-'
MSmits: haha, but once you get going...
LelouchVC2: At least you have a wife
MSmits: he's not complaining about having a wife lol :P
pb4: I did find interesting that there were different approaches to the game at the top
MSmits: yeah, i saw that in eulers stats
MSmits: hey, i Q-learned an AI for NIM
MSmits: so next up, alpha zero :P
pb4: ha ha
pb4: ha ha ha
MSmits: this was just for an AI class I am doing
pb4: Table-based ?
MSmits: need the study credits to become CS teacher
MSmits: it's a dictionary with states, so i guess so
pb4: Q-learning is well-behaved and proven to converge in its table-based form
MSmits: seems pretty obvious that it would converge
MSmits: i also know a bit more about these convolution nn's
MSmits: but i find it hard to imagine training for an entire CG game
MSmits: so many states during a game and only one reward at the end
pb4: If you really want to try, I suggest starting small on an easy game before trying harder games
MSmits: yeah i know
MSmits: I was thinking of doing it over the summer
MSmits: once my teaching is done
pb4: CSB - runner only - no opponent : that's a very good starting point
MSmits: yeah, problem is i dont like csb that much. I would prefer something like oware, not necessarily oware, but some board game
pb4: You'll need some tooling, but there is some existing that you can reuse
struct: connect 4 should be one of the simplest
pb4: Board games bring two more difficulty compared to CSB :
MSmits: discrete input
MSmits: i know this
MSmits: whats the second?
pb4: It's more binary : a single change will make you win or lose
MSmits: thats ok, i can start on tic tac toe as well
pb4: and you have to express the game as a game with final states
MSmits: not sure how that would work
pb4: in CSB you can just say that there are infinitely many CP to take, and the objective is to pick them as fast as possible
pb4: In nim, at some point you don't propagate gamma*best_next_state_reward.
pb4: At some point you've won and you propagate 1
pb4: or you've lost and you propagate 0
pb4: in infiniate CSB with a runner only, you never win or never lose
pb4: All states have the same status, no need to handle "end-game"
pb4: Also there's inherent variety in CSB due to random CP positionning : it's much easier to guarantee that the neural network will have good exploration
pb4: In NIM or board games, it'll get ugly if you don't explore enough or explore too much
pb4: Not saying it's impossible, far from it
MSmits: is this where the drop out thing helps?
MSmits: at least with overtraining
pb4: Dropout is a type of layer in neural networks, it helps with not overfitting the data fed to the network
MSmits: tensorflow seems to have a lot of easy options for this. Should I use that?
pb4: I'm speaking about not feeding always the same info to the NN :)
MSmits: ahh ok
MSmits: well you have data, you split that into test and train data
MSmits: and you can vary the train and test data
pb4: If your NN always trys the same first move in the game, you'll always feed the same value
pb4: Let's say you can play 10 different moves at the beginning
MSmits: so, basically yavalath
pb4: Your NN always tries move 0
pb4: Even if you split between test/train, there's only be move 0 in the train data
MSmits: hmm right
pb4: It's not exactly the same, but think about a MCTS with C = 0
Illedan: Might work with an heuristic yeah
pb4: @Illedan, if I may suggest : try to see how high you go without spells
pb4: It's probably with simpler for smitsimax, and it can go pretty high
Illedan: I do spells as a heuristic on the side now
Illedan: Didn't add them to the search hyet
tutubalin: We found an issue in Van Eck's sequence puzzle https://www.codingame.com/ide/puzzle/van-ecks-sequence
pb4: #12 without spells
MSmits: WINWINWIN they dont use monte carlo though
Illedan: pb4, that tells me my eval is very bad at the moment :P
MSmits: monte carlo is almost always weak
tutubalin: It says that sequence starts from A0 and we need to find An, but according to tests, it uses 1-based indeces. So it starts from A1.
pb4: Illedan : hence my suggestion :D
pb4: This way you'll know when your eval is not the limiting factor :D
Illedan: true :P
MSmits: you need to simulate everything if you want to do it right
pb4: bullet-proof engine ?
Illedan: I sim everything and can play the full game vs myself without getting an off by 1 error
struct: What depth Illedan_
Illedan: 2k playouts
MSmits: in csb i do 7
Illedan: I think I need more sims to go deeper
struct: You prune moves?
MSmits: 2k is very few indeed
Illedan: 8k sims
pb4: Don't bother with more playouts though
pb4: depth 4 is perfectly fine
MSmits: it's a very chaotic game
pb4: at least on my GA, depth doesn't really matter
MSmits: more depth could be meaningless
pb4: just need >= 3
MSmits: same with meanmax right?
Illedan: My meanmax bot has depth 2 :P
pb4: No idea, didn't try smitsi on meanmax
MSmits: i mean the depth thing
MSmits: it's also a low depth game
pb4: probably yes
pb4: We compress it
pb4: for an example
struct: only on contests
struct: multiplayer you can do w/e
eulerscheZahl: i used it for STC in the past but not anymore
eulerscheZahl: at some point CG updated the languages and my submission broke, easier to ensure your bots are still working when you don't submit binaries
eulerscheZahl: and less risk of losing your code
struct: LoGos only during active contest
eulerscheZahl: my bot even uses C#. That tool is for C++
RoboStac: LoGos - my bandas submission is a NN using ~22000 16 bit floats
pb4: AZ ?
pb4: I should go back to making my AZ work on Oware -_-
pb4: still bugs me that it fails so hard....
RoboStac: based on alphazero (the google mcts+nn hybrid)
MSmits: LoGos karlis o's bandas bot is not a nn.
MSmits: you dont need a nn to be competitive is all I'm saying
MSmits: at least not in bandas
struct: csb is the exception for now
MSmits: it seems so, but it's not 100% sure
MSmits: the distance between Pen's non-NN bot and anyone elses was enormous. If Pen could be so much better, maybe something else can also. Something that will compete with the NN's as well
struct: He never said what he was did he?
LoGos: I think I understand his point. He is telling me that while NN sounds powerful that does not mean everyone should use it. Karlis' bot is an example. Yeah? thank you MSmits. I have another question. Do you send your code as binary? I'm just asking because I don't want to waste time optimizing my code while I could use it too
struct: What do you mean optimizing?
struct: Im not sure that binary will increase your performance marginally
struct: I mean that might only increase it marginally*
RoboStac: as long as you have the optimisation pragmas sending as a binary barely makes any difference now
RoboStac: at least for c++
pb4: and in Vindinium's case, it made a large negative difference :/
pb4: Still don't know why
MSmits: I never used send binary
MSmits: i think you need linux for it too
MSmits: I am on windows
Nerchio: why you can't find many top players in multis? they don't resubmit after challenges end?
MSmits: several reasons
MSmits: some contests dont immediately become a multi
MSmits: well they do now
MSmits: but before they didnt
MSmits: also every contest has a different set of players participating, some always join but many dont
MSmits: and the contests span like 8 years or more
MSmits: not all players have been on CG that long or have been here a long time ago and never came back
MSmits: you wont find my bot in ex-contests of 3 years ago probably
MSmits: with some exceptions
Nerchio: makes sense i guess :p
MSmits: it is a bit hard to define a top player btw. For some players its easy to say, but other players are almost always around 50-100 in rank and then suddenly one contest really clicks with them and they get top 10
MSmits: the games are different and different skills may be required
IAmNoob: hi guys!
AntiSquid: Automaton2000 is a girl
Automaton2000: you can just take a look
AntiSquid: ya have a look
AntiSquid: 2014 first multi MSmits
Nerchio: is there any easy way to ctrl+c the output
IAmNoob: i dont think so
IAmNoob: but what are the width and the height of the screen if "Coders Strike Back"?
Nerchio: idk never played it
IAmNoob: i think the height is a little smaller
MSmits: you can go off screen
MSmits: so it doesnt matter
MSmits: not that there is any reason to go off screen ofc
MSmits: but you can
IAmNoob: not in pixels
MSmits: pixels dont matter
IAmNoob: i mean in the measrement
MSmits: positions are doubles
IAmNoob: that they use
MSmits: yes they dont use pixels
MSmits: they use doubles
IAmNoob: i know
IAmNoob: how much is the screen in doubles
MSmits: the viewer is meaningless really, it's just so you can see where your pods are
IAmNoob: i know
MSmits: i dont understand your question IAmNoob
IAmNoob: sry i am unclear
IAmNoob: if i want to give my pod the coordinates of the center
IAmNoob: what should they loock like
MSmits: ohh I see. Not exactly sure, but if you want a "center" you should probably just take the average position of the checkpoints
IAmNoob: and how can i do that
IAmNoob: because at the start i dont know all of the check point
MSmits: add up all the checkpointX and divide by number of checkpoints
IAmNoob: i know how
MSmits: dont overthink it if you're in wood 1
MSmits: or wood 3 i dunno
MSmits: make a simple bot
IAmNoob: i am bronze 2
IAmNoob: so i need:(
MSmits: bronze 2??
MSmits: dont you mean wood?
MSmits: there is no bronze 2
IAmNoob: ok but they arent the same every turn
IAmNoob: now in bronze
MSmits: dont you get all coords in bronze?
RoboStac: no, not till gold
MSmits: thats silly
MSmits: csb really needs some revamping imho. It's the worst tutorial
MSmits: gives the wrong impression about CG
RoboStac: the tutorial bits fine, it just leaves you in a really awkward place
MSmits: sorry, i mean the whole progression to gold
Astrobytes: That's putting it lightly :)
IAmNoob: no problem
IAmNoob: its ok
MSmits: should we just give him the "trick" ?
MSmits: then he gets to a league with coords
IAmNoob: dont want
MSmits: good answer
Astrobytes: I don't usually tbh, I find they learn more by doing it the long way
IAmNoob: i'll make it all aabout my own
MSmits: they do
MSmits: I just dont want them to get frustrated with the changing input
Astrobytes: Yeah I know, it's a silly situation
MSmits: it's just a heuristic to prevent oversteering, you can come up with it yourself
MSmits: so you dont go in circles around checkpoint
IAmNoob: no :(
MSmits: i am not giving the trick
IAmNoob: i dont need any
MSmits: i just describe a problem that you need to fix
IAmNoob: i'll make all about my self
Astrobytes: But let's just say: Don't multiply your velocity by pi
MSmits: ha no, dont do that
AntiSquid: i would just start with a different multi really
AntiSquid: most multis are easier
MSmits: some, not most imho
AntiSquid: all grid based ones (except for UTG OOC) are easier
MSmits: if you know how to write a search
Astrobytes: I think I've seen this conversation 3 days in a row on this chat, maybe 4... :)
AntiSquid: even if you don't
MSmits: not all of them have easier heuristics
MSmits: try it with oware
AntiSquid: oware isn't actually a grid
MSmits: it's a 2 by 6 grid
AntiSquid: play kutulu Willem640
MSmits: kutulu is cool
IAmNoob: it make my code brocken
struct: Wondev Woman is good
MSmits: kutulu teaches you simple searches like bfs
IAmNoob: dfs is easier
IAmNoob: and shorter
MSmits: dfs and bfs are equally hard
AntiSquid: i wouldn't call it grid MSmits , because it's irrelevant how you represent it, you can have a list where you connect the start and the end
MSmits: you're just replacing a stack with a queue
IAmNoob: but is shorter to type
IAmNoob: the dfs
IAmNoob: in c++
AntiSquid: actually you can't move up and down at any point
IAmNoob: is only 5 lines
MSmits: ok AntiSquid, but that makes it even easier than a grid :)
MSmits: cuz its 1 dimensional
MSmits: it's a 1 dimensional grid :)
MSmits: circles are 1D
AntiSquid: point is gird games are easier .
MSmits: the representation and the sim is easier
AntiSquid: circles are 1 D ? in what universe
MSmits: they are
Nerchio: damnit i started meeting people who do the mine trap on 1st column in crystal rush xD
AntiSquid: no the aren't :/
MSmits: surfaces of spheres are 2D
MSmits: circles have no thickness, its a line with the 2 ends connected
MSmits: lines are 1D
AntiSquid: you can only represent a circle in 2d space, otherwise it's a line
MSmits: actually you can represent all points on a given circle with 1 coordinate
MSmits: so it's 1D
MSmits: it's a 1 dimensional curved space
struct: circle is 1d indeed
Astrobytes: You can only travel in 2 directions
Astrobytes: so it's 1D
Astrobytes: But in the 2D space
AntiSquid: if you fold the piece of paper around the line sure you can get a circle but that's not real 1d
MSmits: mathematically it really is 1D, the conceptual circle that is, not the actual paper one, that's 2D
MSmits: or 3D even
IAmNoob: I am a math student and i am sure the circle is 2D
IAmNoob: but is a vector
IAmNoob: which is 1D
IAmNoob: so the both are correct
AntiSquid: represent everything in a 1 dimensional array, he it's 1D !
Astrobytes: Smits explanation is on point
IAmNoob: point is not
IAmNoob: array of points
Astrobytes: It really doesn't need much explaining
AntiSquid: pointless conversation, point was grid games are easier
MSmits: yes they generally are
MSmits: to simulate that is
Astrobytes: Mostly yes
MSmits: to search, not necessarily. Some are hard to find good heuristics for
MSmits: so then you need something like minimax
AntiSquid: physics games can require googling for trigo formulas so eh ...
MSmits: right, difficult sims
AntiSquid: you don't need that much knowledge for a basica grid .
MSmits: i separate sim and search
MSmits: a sim is not an algorithm
MSmits: not a search algorithm
Astrobytes: *may contain algorithms
MSmits: to me the actual writing of a bot is not in the sim, thats just copying the game and it's annoying
MSmits: unless it involves bitboards
Astrobytes: lol MSBits
MSmits: the sim really turns me off from FB, meanmax etc.
MSmits: so i see AntiSquid's point for sure
Astrobytes: Yeah, I'm still putting those off
MSmits: but i see it more as a lot of boring work to do
Astrobytes: It can be tedious
MSmits: thats quick
Astrobytes: Are you now in gold?
MSmits: not sure if thats the trick
Astrobytes: It's not the trick then.
MSmits: you have to somehow get your bot to start steering before it gets to the checkpoint
MSmits: like you would in a real race
Astrobytes: Don't wait for any other bot to do something, take charge of your pods
struct: The trick doesnt beat wood 2 boss I think
struct: but beats wood 1 until gold
MSmits: i dont see why it wouldnt?
MSmits: yes, sick drift
MSmits: let's call it the sick drift trick from now on
Astrobytes: Yes, you're getting the idea
struct: no collisions, and wood 2 turns preety well
struct: and always same map
Astrobytes: Code Sick Drifts
struct: 1vs1 you wont win vs boss
struct: with trick
MSmits: really... too bad i cant test
IAmNoob: Drifts are a way to go
IAmNoob: i am sure
IAmNoob: but sometimes
struct: Ill test give me 3 min
IAmNoob: it gets stuck
MSmits: Willem640 I got to gold with some weird search with vectors
Astrobytes: I did vectors and steering behaviours
Astrobytes: Loads of ways to do it really
MSmits: i wrote 200 lines, but this trick gets you to gold with 10 lines or so... i thought it was annoying =/
Astrobytes: Yeah. I just feel you miss out if you don't do it the long way, as I said before. Seems a shame to waste the experience
MSmits: you know where the checkpoint is, use that
MSmits: and correct somehow for what your pod is doing
IAmNoob: is there a silver too?
IAmNoob: after those
MSmits: bronze silver gold legend
struct: lol the scores on wood 3
struct: are all negative
struct: even boss
Astrobytes: oh Smits beat me
IAmNoob: wood 3 i got top1
MSmits: someone do some wreckin' struct?
darkhorse64: Is there a way to know which puzzle you have done per language ?
Astrobytes: probs all broken bots struct
struct: There is a site I think darkhorse64
MSmits: ow hey darkhorse64, I wanted to share something funny with you. i'll pm, already told it on WC twice
darkhorse64: wrong fitting ?
MSmits: oh you read that
MSmits: cool huh, my bot wa awesome at seed 0 :P
darkhorse64: I was definitely surprised that you did not crush the leadderboard
Astrobytes: still funny :)
MSmits: you mean when i shared my cg bench?
darkhorse64: which site struct
MSmits: yeah you were right it was weird
MSmits: but obvious now
struct: As you can see
struct: the trick gets beaten
struct: By wood 2
Astrobytes: lol nice
MSmits: seems to me the trick is not done right?
AntiSquid: UTG has a really terrible referee, who wrote that code? it's even worse than mine
MSmits: steers too late
MSmits: yeah definitely too late
struct: But the code is right
MSmits: what is yours AntiSquid
struct: If I pass wood 2
MSmits: can pm me the code?
struct: Ill get gold with exact same code
MSmits: yeah ok, but the code may still be wrong :)
AntiSquid: which one?
AntiSquid: was saying UTG referee code is worse than what i can write
MSmits: oh i thought you meant a specific contribution of yours
MSmits: i was just wondering which of yours you thought were bad
MSmits: btw struct is right the map is bad, his code is perfect :)
MSmits: but the trick is the same, prevent oversteering, just dont do it the way you would in bronze or higher
struct: The trick for wood 2
struct: Is just do what they tell ytou
struct: They give you pseudo algorithm
Astrobytes: aim at the cp
IAmNoob: yes :)
Astrobytes: Greetings Lord Zarth of aaX
struct: ZarthaxX you join next contest?
ZarthaxX: hello Lordastro
ZarthaxX: i suppose tobo struct
ZarthaxX: but wonthave much time due to uni
struct: on gold they give you velocity
struct: and all checkpoints at start
reCurse: Tracking velocity is very easy though
MadKnight: but it's not as long as tracking all checkpoints
LoGos: maybe this will help http://cgstats.magusgeek.com/app
MSPK: hey,can u pls share the code
reCurse: It's working
reCurse: Well read the full error
reCurse: "Please note that cgstats can't find your ranking if you are above the first 1000 players."
reCurse: Well the translation kinda sucks
reCurse: It means if your rank is below 1k
reCurse: Or above 1k if you consider number
reCurse: In other words
reCurse: It only works if you're in the top 1000
reCurse: Yeah the translation is bad
reCurse: What data do you want and why?
reCurse: Ah I see. cgstats won't help then.
jacek: uncertainty principle tells you cant have both
jacek: deal with it
reCurse: You'll need to parse the data yourself
reCurse: Optimal control is also useless starting in gold league
reCurse: Good enough control
reCurse: The game completely changes in gold league once you get two pods
reCurse: You'll constantly have one enemy pod blocking yours
reCurse: So the "optimal" trajectory will be extremely hard to compute
reCurse: That's why I'm saying to focus on what's good enough and not overthink it
reCurse: You definitely don't need to parse data to reach gold
AntiSquid: no need to torture yourself
Illedan: Let me know if you are able to do it :D
mark7: solved Mars Lander Ep. 2 with 100% and all 3 achievements now I could sleep :joy:
Astrobytes: good job mark7
mark7: Astrobytes thanks :-)
ZarthaxX: ah so you copypasted, gj
Astrobytes: He's not a copypaster ZarthaxX of ZarthaxXia
mark7: Astrobytes what is he talking about?
Astrobytes: mark7 a lot of people copy/paste solutions from github. A LOT.
Astrobytes: But I remember you trying real hard in the contest, so I think you're not a copy/paster
mark7: mark7 no I didnt copy/pasted because if doesnt add any value knowledge to me
mark7: working hard, sharpening brains make you smarter :-)
Astrobytes: It certainly does. So does studying for your exams btw :P
mark7: Astrobytes :joy:
AntiSquid: aren't schools still closed?
reCurse: Reopening soon here
reCurse: You mean to get into gold or get into silver?
reCurse: You don't need shields
reCurse: But it might help
EnragedOxygen: Ramming is wonky without knowing enemies velocity. And trying to defend from accidental bumps slowls me down more times than it helps. Am i missing something aka a better raming algoritm?
reCurse: Hmm I forgot the limited inputs
reCurse: Yeah I guess you could totally ignore the shield
AntiSquid: best tutorial for new users
reCurse: What would you suggest
Astrobytes: Perhaps introducing a (skippable of course) tutorial for a couple of other CG multis would work. Then again, it's probably too much work to update the old ones.
AntiSquid: to not recommend the game
reCurse: Great suggestion
reCurse: How about an actual game you think would be better suited?
AntiSquid: even WW would be better
Astrobytes: The only problem with CSB being the tutorial game is the change in Gold tbh.
Astrobytes: imo, before I get flamed
MSmits: the problem with csb is that the input is inconsistent and people always come here and ask why their bot doesn't work. Then we have to tell them to reset their code
MSmits: something you dont have to do with any other multi i know
reCurse: I'd say you stop being a new user once you reach gold.
Astrobytes: Depends how you get to gold tho
MSmits: -pi vel?
Astrobytes: just pi * vel, noob :P
MSmits: ow right, for the circles
AntiSquid: promoting from silver is full on kick in balls golden experience
Astrobytes: But yeah, a lot of really green users do the MK formula and then just stop
MSmits: they should really be referred to a simple board game
Astrobytes: One size doesn't fit all
AntiSquid: just don't promote csb . simple :D
MSmits: I mean for simplicity. CSB is a dead end in gold
MSmits: for new users
AntiSquid: it has 90k dead accounts or so
AntiSquid: maybe it's not a good idea to promote it
Astrobytes: So there should be a couple of tutorials I think, of varying types of multi
AntiSquid: well you can refer new users to codingame videos
AntiSquid: on youtube channel
reCurse: You keep saying to not promote it but fail at suggesting better
Astrobytes: It won't happen anyway so we're debating pointlessly, but in the Ideal World...
reCurse: WW is not easy at all to start with
AntiSquid: i said even WW would be better, not that it's great
MSmits: uttt has the regular uttt board to start with
MSmits: think its a pretty big jump to the 9x9, but its an ok start
MSmits: tron works ok
Astrobytes: Someone suggested just having a regular TTT for practising
reCurse: You're comparing something that requires search
reCurse: With something that requires intuition
reCurse: CSB is just fine, the leagues are unfortunate but can't be fixed
Astrobytes: Yes, fair point.
MSmits: I know, but that's my point, once you get gold in csb you need to search to get any further.... if you're going to learn how to search, maybe csb is not the best place to start
reCurse: Once you reach the top at any game you'll need to search
reCurse: Your point?
MSmits: my point is, suggest other games that they can learn the skills in more easily
AntiSquid: uhm you need to track shit in csb silver and bronze
reCurse: Yes the leagues are unfortunate
AntiSquid: you can reach silver / bronze just as easy in most multis
reCurse: If your onboarding requires search it's a shitty onboarding
MSmits: yeah dont use search in onboarding
MSmits: use q-learning
AntiSquid: if your onboarding requires more trigo than programming it's a shitty one @_@
reCurse: lol ok
Astrobytes: Perhaps the CG redesign with the learning track thing will help with learning particular algos/techniques
reCurse: Conversation with no purpose
AntiSquid: lol ok back to you
MSmits: yeah thats really what i was trying to get at. The transition from heuristics to search could use some streamlining on CG.
Astrobytes: Hopefully that's considered in the redesign. When it comes. If it comes.
Astrobytes: And that leaderboard better not be indicative of the coming desing
MSmits: which one
reCurse: Prepare to be mostly disappointed
Astrobytes: The squished up one
MSmits: is it still up?
MSmits: i dont see it
Astrobytes: Yeah reCurse, I'm not optimistic
Astrobytes: It seems to have been reverted (being Friday, always expect CG updates) but it's burned into my mind now
reCurse: Uh? Do I need to refresh
reCurse: Looks like it
reCurse: Yay SPA
reCurse: Single page application
reCurse: Nothing refreshes for real
reCurse: Until you do
reCurse: I think it's time to submit csb
Astrobytes: Right, so they've been seeing us moaning. So lets moan more.
reCurse: Next step is to do machine learning to recognize wood bosses and have my personal DDoS tool
reCurse: Don't ruin the fun and ask why
reCurse: (joking in case)
Astrobytes: I can't see the CSB submit issue being resolved anytime soon
reCurse: I can guarantee doing that during a contest will get the issue addressed asap
Astrobytes: It happened already tho' no? Was it UTG?
reCurse: Yeah sort of
reCurse: Would be 10x worse if deliberate
Astrobytes: Then we need coordinated submits :P
MSmits: They should at least have a button to stop all multiplayer activity in case of short contests
Astrobytes: The multi AZ5
MSmits: ah chernobyl
Astrobytes: SCRAM/AZ5 isn't just limited to Chernobyl ;)
Astrobytes: Killswitch, there.
Astrobytes: Big Red Button on Thibaud's desktop.
reCurse: Which creates a JIRA and assigns it to someone to do the killswitch
Astrobytes: And 8 months later the issue appears as resolved in the bug channel on Discord
reCurse: No for emergencies they are quite reactive
reCurse: I just don't think Thibaud has the rights :P
Astrobytes: True. lol, they should give him more power
Astrobytes: Grrr. There's a bug in Tulips and Daisies. When I make changes I lose. Where do I report this.
Astrobytes: It's clearly a conspiracy.
reCurse: Your brain is staging a revolt?
Astrobytes: I just miss our "the site is bugged my code doesn't work" users, so I'm projecting
Astrobytes: That and I fucked up my eval again
inoryy: first time it happened over a year ago, somebody boosted like 500 accounts in CSB and leaderboards were halted for almost a day
Astrobytes: Can you not do that please usaykurmoto
Astrobytes: I think I remember that
Astrobytes: how's it going inoryy?
inoryy: aside from going crazy over sitting at home for 2 months? good good :)
reCurse: You and me both
Astrobytes: Tell me about it.
inoryy: more or less settled in, finished a mini-project today
inoryy: excited to start the next one on monday
Astrobytes: Nice. Anything interesting?
Astrobytes: *particularly interesting
inoryy: I think so :) But probably won't be public for awhile
AntiSquid: i am feeling fine, just taking a walk whenever it's sunny, shop something, go home, coworkers sound stressed though
Astrobytes: hehe :spy:
AntiSquid: inoryy when is there going to be a reveal about anything ML related from deepmind?
Astrobytes: I've got loads of issues, mostly related to my Grandad, since most social care has evaporated
Astrobytes: When they have finally taken over the world AntiSquid :P (joking)
AntiSquid: how come? what happened to the social care?
AntiSquid: what changed
inoryy: AntiSquid what do you mean? we have publications coming out almost daily, sometimes with detailed blog posts :)
AntiSquid: then i am bad at finding them
AntiSquid: forgot that site
AntiSquid: but is there anything major? like another game playing model ?
Astrobytes: AntiSquid Social services (adult) washed their hands of him at the start of the pandemic. When I was isolating, the only way I could get food to him was via someone I haven't spoken to for 20+ years that lived nearby. Medical-wise, there's no podiatry, no diabetes clinic, no dermatology clinic. All we have is the care workers 4x a day
AntiSquid: you know stuff that doesn't happen all the time
inoryy: these two are quite major: https://deepmind.com/blog/article/Towards-understanding-glasses-with-graph-neural-networks
AntiSquid: so pink background = major, noted
Astrobytes: Nice, the glass one looks very interesting
reCurse: Personally the development I've been the most impressed by is those imitating extremely complex physics at a fraction of the cost
inoryy: yeah, it significantly outperforms current SOTA and has the potential to be beneficial in many science fields
reCurse: That's just mind boggling to me
inoryy: yeah that's the first link reCurse :)
AntiSquid: don't remember seeing anything about graph neural networks before
AntiSquid: might be old concept but new to me
reCurse: I was thinking of something else
inoryy: oh I think I know which one you mean
inoryy: https://arxiv.org/abs/2002.09405 ?
reCurse: Yeah and there was another with an impressive video with floating capes and stuff
AntiSquid: quick summary for graph netwroks please?
reCurse: Trying to find it...
reCurse: I think it's this one http://cim.mcgill.ca/~derek/files/Deep-Cloth-paper.pdf
reCurse: Or some followup
Astrobytes: I'm a fan of the biological/biochemical applications too, protein-folding etc. Revolutionary.
reCurse: Having a fixed compute for complex animation or physics is just insane
inoryy: AntiSquid GNNs approximate graph structure via NN, typically through layers that work well with local spatial connectivity, e.g. convolutions
Astrobytes: You have the link to the video for that reCurse?
reCurse: I just noticed ubisoft contributed to that paper
reCurse: So it's possible I'm referring to stop I actually saw internally lol
Astrobytes: Seems rather impressive
reCurse: Makes no sense to me how well it works
AntiSquid: so convolutional NNs are part of GNNs ? i simply don't remember this term nor the visualization for it
reCurse: Astrobytes: https://www.youtube.com/watch?v=yjEvV86byxg
reCurse: Not the best I've seen but it's something
inoryy: conv is one simple approach to do GNNs but the novel results like glassy dynamics use specialized approaches with embeddings on vertex and edge level
AntiSquid: i clicked a link in that article, shows a paper back from 2018 :/
Astrobytes: The GPU/CPU stats... lol
Astrobytes: Very nice
Astrobytes: *usage stats
reCurse: Yeah it's ridiculous
reCurse: When you compare to how much the ground truth costs
Astrobytes: Indeed. Game changer. If you'll pardon the pun