Chat:World/2021-07-14

asdfg47: :thumbsup:

derjack: good morning

dbdr: \o

FFly: hello world

derjack: :upside_down:

FFly: :grin:

JormaW: can someone help me publish a puzzle here? Having trouble running the tests

derjack: what puzzle

JormaW: an easy puzzle ive got

derjack: what trouble

JormaW: You get an input between 2 and 26 and according to that you should output a square of letters so the box always has one A in the middle of it, the next layer will be B:s all around the A, the next layer will be C:s all around the B:s, and so forth all the way up to 26.

BlaiseEbuth: And what do you want to do with that ?

JormaW: Output the letter square on the screen.. print it?

BlaiseEbuth: Yeah

JormaW: Thought it was a pretty hard puzzle for me to do and I've gotten 4 easy puzzles done here so I thought it could fit in the easy category

BlaiseEbuth: I don't get if this is a puzzle you're solving or creating... Need more coffee

JormaW: Im trying to create it but couldnt get the tests to go through

JormaW: Coffee sounds good though

KiwiTae: there is a clash of code like this

JormaW: How can i find those?

BlaiseEbuth: Yeah I think this is more fitted for a clash than a puzzle

JormaW: I mean i dont want to make a duplicate of the puzzle if there already is one but I'm not really sure how to find them either.. Is there an easy clash category that I could browse as well?

BlaiseEbuth: Not directly sadly. But you can try with https://eulerschezahl.herokuapp.com/codingame/puzzles/

JormaW: oh wow

BlaiseEbuth: And hi eZ ! Sorry for the ping :p

NavLol: sup

dbdr: \o Marchete

Marchete: o/

dbdr: you have some useless inputs in oware

Marchete: it seems a better bot, isn't it?

dbdr: seems so

Marchete: what useless?

dbdr: scores >= 25 are never used

dbdr: since it's a win

Marchete: yeah

Marchete: I saw it

dbdr: why did you pick <27?

Marchete: it should be 26

Marchete: 0..25

dbdr: 25 excluded

Marchete: I changed to SIZE_CELL_NN=28 SIZE_SCORE_NN=26

Marchete: but I haven't seen any improvement

Marchete: I think 28 is too much

Marchete: 25 excluded?

Marchete: well

Marchete: it's true

Marchete: once you hit 25 it's a win

Marchete: so 0..24

dbdr: also score 1 is impossible

dbdr: so it can be packed on 3 avx regs :)

derjack: :o

Marchete: how 3?

dbdr: 3*8 = 24

BlaiseEbuth: lier

dbdr: 0 and 2..24

derjack: do you have limit on max seeds per house? i assume 31 the max

derjack: maybe 23 seeds max would be sufficient

dbdr: yes, limit is 23 in stock cgzero

derjack: oh?

Marchete: stock :D

derjack: vanilla

dbdr: vanilla if you prefer

Marchete: I just put some value there, as a guess

dbdr: arf

Marchete: it's not like anything well thought

derjack: so maybe you could benefit more from more seeds :v

Marchete: "640KB should be enough"

dbdr: again 0..23 is good as a multiple of 8

dbdr: I haven't seen huge stacks (kroos?)

Marchete: but I no longer need an "input" for one hot

Marchete: I directly do the first layer output

dbdr: ah, you optimized that?

Marchete: I mean

Marchete: reset() -> set first layer to bias

dbdr: yes, you can just index

Marchete: then setbit(N)

Marchete: add to that layer weights[N]

Marchete: then once you have that layer

JormaW: Is there someone here who could help me with a small flask/html/sql project problem?

Marchete: ReLU(intermediate,output)

Marchete: I don't know how much that improves

Marchete: not enough considering you have +1.6pts over me

dbdr: you mean you don't measure? :scream: ;)

Marchete: measure what

dbdr: speed

dbdr: ah you mean TS points?

Marchete: leaderboard points

dbdr: right

dbdr: I hope you know how much speed it gives

Marchete: you need some better winrates against smits, recurs_e and you can be 2nd

dbdr: yes, pretty good against top 3

Marchete: nope, I think "a bit"

Marchete: 0/1/11 vs Robo is a huge win right now

dbdr: bad against 6-8

dbdr: I implemented the weighted random

dbdr: without temperature

Marchete: agad_e as p2 is weird, he doesn't play the normal move

dbdr: gave this bot in one gen. not sure if this is a direct consequence or just lucky

Marchete: I implemented it, but I fear it polluted the samples

coolbandit99: hi

coolbandit99: :grinning:

derjack: top is pretty much rps

derjack: ohai

dbdr: why polluted, aren't they excluded?

Marchete: it depends on how you implemented it

Marchete: why they should be?

dbdr: I was thinking better (weighted) random moves should improve a bit the samples for the value predictions, since the bot is stronger. might be negligible

Marchete: did*

Marchete: how did you implemented it?

dbdr: it's what you mentioned in the final words

Marchete: I know

dbdr: those samples are excluded by your implementation

Marchete: all of them?

dbdr: you even have a comment about it:

 //I think this doesn't compromise sample quality, because the "offending" selection won't appear in samples.

Marchete: where, in my code?

dbdr: yes

Marchete: let me rephrase

Marchete: I added some random selection, with rnd.NextInt(1000) < 15

Marchete: so it only happens in 1.5% of samples

dbdr: actually, the samples have the visit distribution, not the selection, right? so why excluded it at all?

Marchete: and I marked them as dontGenerateReplay

dbdr: yes

Marchete: but the point is that in original papers

Marchete: they don't say anything about that

Marchete: they just say "weighted random selection"

Marchete: so are you doing a weighted with 1.5% chance

Marchete: or always?

dbdr: you mean you fear it would pollute if you didn't have dontGenerateReplay, right?

dbdr: 1.5%

Marchete: exactly

dbdr: ok, I misunderstood

Marchete: I initially added without the 1.5%

Marchete: and without dontGenerateReplay

dbdr: so you improved alphazero. or discovered an aspect they failed to document

Marchete: so it was adding a lot of noise

Marchete: I mean

Marchete: if you generate 5000k games

Marchete: it's good

dbdr: but the samples contain the selection?

Marchete: because white noise will be cancelled

dbdr: I thought it had the visit counts

Marchete: in the original paper they don't talk about how that random selection affects sample quality

Marchete: I think it affects

Marchete: if you random move and lose, you shouldn't learn that part

Marchete: what I was trying are two things

Marchete: 30% chance of random selection

Marchete: then set:

Marchete: if (tmpValue < 0.87*bestValue) { dontGenerateReplay = true; }

Marchete: mark as bad if the selection differs a lot from best value

Marchete: maybe it's better 0.95 or something higher

Marchete: I was also trying that in the replay buffer

Marchete: if I "win" with a random move, try to use it as a good sample

dbdr: you don't store the sample, but the parent is still affected, right?

Marchete: but I think I have sign problems

Marchete: no

Marchete: I backpropagate the boolean

dbdr: ah ok, so only children are saved, to add variety

Marchete: https://github.com/marchete/CGZero/blob/master/src/CGZero.cpp#L2537-L2551

Marchete: yeah

Marchete: you can't sample a loss if later moves were random

Marchete: but you get more variety on depth

Marchete: I think that's "better" than original

Marchete: because original is more brute force

Marchete: do zillion of games so all that noise will cancel out

dbdr: assuming original (paper?) saves all replays?

Marchete: 500k games

Marchete: If I recall

Marchete: in oware that would mean around 80M positions

Marchete: my changes (with sample deduplication and that) saves a lot of time and space

Marchete: and it doesn't need days to train to a good position

Marchete: I don't know your design

Marchete: but I reach top10 like in 2hrs

Marchete: "reach" as winning top8-top10 guys

Marchete: in just some IDE tests

dbdr: I'm still running your code, just tweaking and playing with it to get some understanding

dbdr: so my design is your design ;)

Marchete: you understand better than me :) +1.6 points

martinpapa69: hi

Marchete: I'm retraining with different ideas, but I can't beat my current submitted bot

derjack: yeah, deciding about how to and what to explore is quite important

dbdr: or it's a fluke, against most players there are a few matches that repeat over and over

Marchete: in fact yes

dbdr: szia martinpapa69

derjack: it sky rocketed when i used softmax with variable temp thingy to decide what moves to use

Marchete: most players are like deterministic

Marchete: variable temp? where?

Marchete: and what game?

derjack: in my own framework, during training

dbdr: Winrate 31.9444 23.4848 I wish I understood what causes such dramatic regressions. does not look right to me

martinpapa69: I too finally sucessfully run your train script with a small NN. compressed with the scripts you added to the playground. but when i run the code in ide, i get a simple:timeout. :D any idea why ?

derjack: keep in mind i only use value

derjack: martinpapa69 any error?

martinpapa69: Game Summary: martinpapa69 timeout!

Marchete: NN weights are called w32?

Marchete: best.w32

martinpapa69: yes

Marchete: probably incorrect binary

dbdr: nothing on console?

dbdr: even the decoder should print something, right?

Marchete: first thing on console is:

Marchete: CODINGAME using Neural Network 0.0900959

Marchete: CODINGAME is even before reading a file

martinpapa69: so i guess, for me, the binary doesnt even start

Marchete: how did you compile it?

martinpapa69: copy pasted the 3 script lines

martinpapa69: i changed "$AI_Name" to CGZero

Marchete: it should work

martinpapa69: not sure if its correct

martinpapa69: okay, i guess i have some debug-homework then before i send it to the arena :D

Marchete: the problem is that you don't have avx2 in your CPU

Marchete: so you can't just test it

martinpapa69: i rewrote the code to awx only

Marchete: ahh

Marchete: then try locally

Marchete: create a net dotnet project

Marchete: add the cs file

Marchete: and dotnet run

Marchete: it should decompress and run the code

martinpapa69: ye good idea, maybe i get some error messages

Marchete: I usually did that to "recover" submitted weights, when I can't recall what generation was

Marchete: if not

Marchete: use the original NN_Mokka.h, and try to compile

Marchete: even if you don't have avx2 it should compile to that target

Marchete: your custom avx it's for your PC

Marchete: not for CG

Marchete: CG has avx2

dbdr: beating robo as p2: https://www.codingame.com/replay/570094556

Marchete: p2 is really hard to win

Marchete: if I ever with Robo is with p1

Marchete: (#UNAUTHORIZED): "You are not authorised to view replay 570094556".

derjack: huh

Marchete: let me in, let me iiiin!

derjack: a mere mortals cant see the robo lose

dbdr: strange

dbdr: this is an IDE match

RoboStac: it's ok, I'm happy if there is no proof of me losing

Marchete: you can repeat the win?

Marchete: NN are very deterministic, at least mine :D

Marchete: can you*

martinpapa69: it looks like there is something wrong with my upx Marchete. after i run the the upx script on the cgzero executable, and try to run the output exectutable it wont print out the welcome message(using Neural Network 0.0255805), or anything

Marchete: then it's wrong, yes

Marchete: always test the binary, if it says using Neural Network

Marchete: it's OK

dbdr: second match as p2 vs robo was a draw

Marchete: that's a win for me too

Marchete: p2 draw == win

dbdr: yours is actually trying to be less deterministic with the bestMoves vector

Marchete: it's barely noticeable

Marchete: also I tried the jace_k way of mixing mean Score and log(visits)

Marchete: maybe it's not enough

Marchete: it seems to me that visit counts are very uneven

dbdr: I think this helps when a good move is discovered late

Marchete: like 1 move taking 95% of visit budget

dbdr: that would mean exploration is too low, possibly?

dbdr: though it can be good for tree reuse

Marchete: in theory it should auto adjust

dbdr: how?

Marchete: dunno :D

dbdr: if exploration is too low it will just exploit

Marchete: policy?

Marchete: I don't know how it solves the case:

derjack: no policy - no probelsm :innocent:

derjack: ~

PepitoPepe: y

Marchete: at some point it thinks move0 has policy 0.98 because it seemed good

Marchete: how to change to move1 as best?

Marchete: it seems hard

Marchete: I was trying to use a higher cpuct on train

Marchete: to give more exploration in learning

Marchete: but I'm clueless (as in most hyperparameters)

martinpapa69: removing the "--ultra-brute" form the upx command solved my problem

Marchete: good

Marchete: then pack and test the bot

dbdr: if move1 is better than the policy thinks, the sample will learn that, even if it does not select it

martinpapa69: i sent it already

dbdr: then next generation it will try it even more

martinpapa69: looks like it wil lend up in mid 10's

RoboStac: I found I got really extreme policies when I had too many mcts iterations

Marchete: how much training?

dbdr: so it might need a few gens to do the switch, if it was not close

martinpapa69: I trained it for like an hour, and my network is rly small atm

Marchete: 1 hour?

Marchete: noice

Marchete: I think my pipeline is fast to train to a good level

derjack: 51% winrate vs random :tada:

Marchete: no no no

Marchete: 46% vs random at gen3!

dbdr: derjack lol

Marchete: :rofl:

martinpapa69: ye its surprisingly fast. every other framework i checked suggests days of training

PepitoPepe: no no no

PepitoPepe: 45% vs random at gen3

PepitoPepe: !

Marchete: yeah, like 3 days on GPU

Marchete: NN models of 3MB size

RoboStac: most online frameworks are for massively deep convolution networks which do take a lot longer

Marchete: that's true

dbdr: maybe other frameworks replicate alphazero sizes, which does not have the CG size and speed constraints. so you need a lot more training

Marchete: but on python

dbdr: that too

RoboStac: oware seems to learn fairly quickly too

RoboStac: my c4 took a lot longer to stabilise

dbdr: c4 with convolutions, right?

RoboStac: nah

dbdr: input of oware is small

RoboStac: couldn't get the performance to work for that

Marchete: it's a pity C4 goes better with dense

Marchete: I wanted to try CNN

Marchete: c4 inputs are smaller

derjack: you can still do exercise

derjack: CNN is just slower

RoboStac: it can still get top 10

dbdr: ah yes, more squares on c4, but only 3 possible states

dbdr: oware has seed counts

Marchete: :lifter:

Marchete: only 1CNN isn't that slower

Marchete: I "converted" it to Dense

martinpapa69: i think the main problem with my oware-alphazero implementations was, that is gave the seed counts directly as an input, not in one-hot format

derjack: but 63 squares in c4, and 14 things in oware

derjack: yeah, who gives inputs directly to NN...

Marchete: each input bit in CNN affect X outputs with W

Marchete: so I recreated these W as weights[InputBit]

RoboStac: direct inputss worked fine for me for a while :(

Marchete: until when?

RoboStac: just before last contest when jacek convinced me to try one hot

derjack: he was in top4 anyway

Marchete: ahh

Marchete: i think one-hot are good for CPU constraints

Marchete: "more info" with just 1 set

Marchete: value_loss: 0.0641

Marchete: if I weight more the value loss

Marchete: like it doesn't miss

Marchete: but policy_loss: 0.7716 :D

dbdr: value_loss: 0.1603 - policy_loss: 0.3325

dbdr: and I kept the 1-1 weights

dbdr: I think you broke something Marchete

Marchete: 2-1 weights

dbdr: ah maybe it's that then

Marchete: I'm just testing dumb stuff

dbdr: try 1-2 ;)

Marchete: I don't know what's better

dbdr: 1-sqrt(2), has better mathematical properties

dbdr: *bullshit

Marchete: why resubmit martinpapa69?

dbdr: while (!first) resubmit();

martinpapa69: ^ :D

Marchete: have you set your cpuct for submit?

martinpapa69: nah

martinpapa69: should i modify it before submit ?

Marchete: you can do it on IDE

Marchete: ;./CGZero");

Marchete: ;./CGZero cpuct 0.0 cpuct");

Marchete: for example

Marchete: at some point of debug text you should see it:

Marchete: "parent_F: CPUCTX.X"

Marchete: just to test on arena

Marchete: I find sometimes best parameters training != arena

Marchete: well, 13th

Marchete: is not bad

martinpapa69: ye i trained the nn for like 1 more hour, wanted to try the most recent one

derjack: already better than cosmobytes

Astrobytes: shut it pony boy :P

Marchete: beat that Astrobyte_s guy :P

Astrobytes: What is that now, 8 or 9 NNs?

dbdr: top 9 - MSmits + martinpapa69?

dbdr: so 9?

Astrobytes: oh, and Ille, he's further down but it's NN

dbdr: shame

derjack: pb4 not NN?

Astrobytes: Hm, you may be right there

RoboStac: I think there might be a few lower down ones too

dbdr: pb4 is suspicious in python

RoboStac: quite a few people were asking me questions and trying to get something working when I first did it

Astrobytes: Tiramo n had one but think it's still in the lower league

dbdr: first other non-compiled language is #42 Zorg1

dbdr: and that's a minimax IIRC

derjack: :older_man:

derjack: i wonder what remi uses https://twitter.com/Remi_Coulom/status/1370475619998392328

Astrobytes: Surprised he hasn't tried Oware

dbdr: different variant?

derjack: there are 3 games i think

Astrobytes: Plenty mancala variants to choose from

Marchete: Smits it's "The indomitable village" from Asterix :rofl:

dbdr: excepts dominated ;)

Marchete: on top10 he is holding

dbdr: yeah

Marchete: and imo he can wreck most NN's with his MetaMCTS

Marchete: and books

Marchete: he just doesn't want to

Astrobytes: Yes, he disabled the booking to keep the peace, so to speak

dbdr: //Play as p0

stopwatch.Start(45 * 1000);

dbdr: stopwatch is shared by all threads

dbdr: it is also unused in pitplay.

dbdr: but that's confusing

dbdr: Marchete

Marchete: yeah, it would create a race condition

Marchete: should be removed from all pitplay and selfplay

dbdr: or keep searching indefinitely :)

Marchete: btw adding a #define RELEASE_MODE and conditional removing most training part

dbdr: no, just longer

Marchete: the binary size is around 30KB when submitting to CG

dbdr: yeah, wasted space :)

Marchete: at some point in the past I had some issues with Random class and threads

Marchete: like using the same object for all threads wasn't a good idea

Marchete: similar to my stopwatch

Marchete: a bad implementation

Marchete: I should define one instance in each worker, pass it as parameter, etc etc

dbdr: in rust thread unsafety is a compilation error, that's good for peace of mind :)

Marchete: indeed

Marchete: c++ is more "oh! you want to add to a vector and at the size time clear it? cool! let's do it and see what happens!"

dbdr: and keep a reference from inside the vector too

dbdr: :upside_down:

Marchete: the worst thing is that C++ tends to invalidate references when adding stuff to a vector

Marchete: sure chaos

dbdr: yeah, as soon as it reallocates

Marchete: like always :D

Marchete: happened to me a lot on NumberShifting code

Marchete: until I found the bug

dbdr: well, growth is usually exponential, like double the size

dbdr: so many pushes will not reallocate

martinpapa69: u can solve it with reservere, when you know the max size

martinpapa69: *reserve

dbdr: which is worse, some bugs will happen only occasionally

dbdr: reserve is good for performance too

derjack: welp, like with NNs, threads will work most of time even if they have serious bug

dbdr: welcome to top10, martinpapa69 :)

martinpapa69: hi :D I wonder how many of marchete's framework will run on hte next contest

Marchete: Zero

martinpapa69: ?

Marchete: CGZero

martinpapa69: noo :(

Marchete: I mean, it's feasible

Astrobytes: *if the next contest is a multi

Marchete: but you need to understand, tweak/recreate

dbdr: next contest will be a golf

Marchete: my framework should be a reference

Marchete: as everybody code in its own way

martinpapa69: dont think its hard to modify the code to run on single player games, like CvZ

Marchete: 10 out of top11 in oware are NN :D

dbdr: to golf, though? the weights take too much space

Marchete: reminds me of GA in CSB

dbdr: oldjohn is NN?

Marchete: lol, it rained since that...

Marchete: I guess, C#?

dbdr: ah. some people actually use C# tho ;)

Marchete: he wasn't C# until recently

dbdr: ok

Astrobytes: truew

Astrobytes: *true

Marchete: the dotnet thing is cumbersome

Astrobytes: Normally uses C++ iirc

derjack: i wonder if the framework were in D, we would see more D people

LlaeSevero: Good Morning Good people

dbdr: should not be hard to write a D packer

BrucevLee: HI kind of new hey

BrucevLee: how can we learn the basic or practice to code

derjack: try onboarding, the descent or temperatures puzzle

BrucevLee: ok thanks

LachlanB: I think im getting a graphical glitch on the AI bot racing, it seems none of the bots move, yet I still win the race. is this a known bug?

derjack: can you share the replay

LachlanB: the replay is fine when I visit the share URL, just when its in the IDE it does not move a frame

dbdr: LachlanB try reloading the IDE webpage

LachlanB: Fixed it, thanks guys. I shouldve restarted as troubleshoot step #1...

dbdr: https://www.youtube.com/watch?v=5UT8RkSmN4k

PanGami: :scream:

LachlanB: Hahaha, oopsie

derjack: :upside_down:

Marchete: how do you improve that fast martinpapa69?

derjack: cloud computing

Marchete: (╯°□°）╯︵ ┻━┻

Marchete: ragequit!

martinpapa69: haha, no i train on my laptop. I added 1-1 layers to the value and policy head, and trained a bit. now im adding 1 more layer to the common NN

derjack: whats 1-1 layer

martinpapa69: 1 layer to value, 1 layer to policy

martinpapa69: and i use adam optimizer instead of sgd. i think it converges much quicker

Marchete: really?

Marchete: I've always heard SGD, and momentum

Marchete: I mean at some early time I tried Adam, but it was before I get it working

dbdr: RL seems like cargo cult programming taken to the next level :D

martinpapa69: it seems to be faster... I used to work with nn-s back than and came to the conclusion, that most of the time you dont have to think about optimizer, just use ADAM :D

martinpapa69: it was years ago tho, dont know if there is some better-adam since than

derjack: i think in AG(Z?) paper they used SGD + momentum, and didnt used other fancy optimizers for some reason

derjack: welp, not that training the NN is big part of learning

dbdr: http://chat.codingame.com/pastebin/c7904d10-3111-4a28-a990-a3034002f15e

martinpapa69: :D

Marchete: in fact in many papers they also said momentum=0.9

Marchete: like it's SGD yes or yes

Astrobytes: The lack of Adamski disspoints me.

Astrobytes: *disappoints

Marchete: the good thing about using commercial ML frameworks is that you can test all this in no time, bug free

derjack: https://datascience.stackexchange.com/questions/30344/why-not-always-use-the-adam-optimization-technique/30347

Marchete: pulling a complete ML framework by oneself it's like creating Linux from scratch...

dbdr: recurs@ did it didn't he?

Marchete: indeed, and probably CSB guys too

derjack: and i did, but i only did basic nn stuff

Marchete: it's great for learning, but I just have no idea where to start...

derjack: that could be also why i stick to value only

dbdr: Marchete: you started, you have the inference

dbdr: I guess the next step would be the backpropagation

Marchete: lel

Marchete: no I'm fine, thanks

Marchete: for me ML is just a tool

dbdr: "You reached the limit..."

dbdr: guess it's time for a break

derjack: hm?

Marchete: stop pressing that button :D

dbdr: it's not me it's my tool :)

Marchete: what tool?

dbdr: cgbenchmark-like

derjack: use the alt

Marchete: I thought it was the play my code

dbdr: that's what the tool does, ide matches

Marchete: CG lowered a lot the matches count

dbdr: yeah'

dbdr: also replay lifetime, recently

Marchete: they should add a delete button for history :)

dbdr: they could probably reduce replay size by 90% easily

dbdr: most of these things look very bloated

struct: sdk replays are huge if you use gem

dbdr: gem?

struct: graphic entity module

struct: recurs e doesnt use it on chess

struct: one of the reason his replays are so small

dbdr: nice

eulerscheZahl: you can also do a mix of graphics module and your own. Or just implement your own serialization and still feed it into the SDK module in the end

eulerscheZahl: 38 frame replay of Dice Duel https://eulerschezahl.herokuapp.com/codingame/replays/analyze/?id=568229737 still some redundancy that we can't get rid of

Marchete: near 7th place, csip

martinpapa69: ye, but its losing its momentum, sp o think if i want to improve further, more training wont be enough

Marchete: how big it is?

Marchete: maybe space is limited

Astrobytes: does renard have NN also? Recent submission

Marchete: it forgets to learn new things

martinpapa69: my binary is 80kb-ish

Marchete: you can add #define RELEASE_MODE to cut a lot of code

Marchete: can be reduced to like 32KB

Marchete: I do conditional compiling

martinpapa69: dunno if the network size is the problem here, but ye ill try that

Marchete: adding -D RELEASE_MODE

Marchete: to enable or disable it

Marchete: maybe you can do a makefile or smth

Marchete: when I asked big, I was refering to weights file

Marchete: 1st layer is usually faster to calc than later

Marchete: as you have a lot of zeros

martinpapa69: the w32 file itself is 83k

Marchete: ok not that small

martinpapa69: ye, i think the netowrk size should be fine, i have to look the hyperparameters now

Marchete: I can't help much more, I also hit a ceiling

Marchete: the bot seems very sensitive to rollout count

Marchete: 10k sims worked better than 5k sims

Marchete: about hyperparameter, no idea :D

Marchete: I don't know if it's better to use PROPAGATE_BASE=1.0 (i.e. like Alphazero does), or learning rate, or pretty much anything

Marchete: and ofc the damn cpuct

Marchete: I think I know how leela chess do

Marchete: it's related to visitcount

Marchete: but that's visit count doesn't change that much

martinpapa69: no idea. i've only tried changing PROPAGATE_BASE, but it didn't seem to have big impact

Marchete: but I don't know if cpuct must be 1.0, 1.5, 2.0,2.5 or 3.0

Marchete: who knows

Marchete: also I haven't tested much about FPU (first play urgency), they said -1.0, others -parentValue

Marchete: I also don't know

struct: :)

struct: i wonder if i finally should port arimaa

struct: I want to know if NNs do well there

struct: normal searches dont

dbdr: ¯\_(ツ)_/¯

struct: if i remember its like 4 moves per turn

dbdr: sounds easier for search, low branching

struct: maybe

struct: i meant that you do 4 moves per turn

struct: not 4 moves available

dbdr: ah, got it

AllYourTrees: Marchete i think noise weight, cpuct, and num simulations are all interconnected. iirc the goal is for the search to overcome noise, so too low num simulations or too high noise will make it so search cannot overcome noise

AllYourTrees: err maybe goal is the wrong choice of words - the noise should be at a level where the search can overcome it if it doesnt benefit the probability?

Marchete: noise on a big amount of samples should be cancelled

Marchete: but I don't have 5000 TPU years to test

AllYourTrees: ya haha

Marchete: or crazy numbers like Google

Marchete: I think dirichlet noise is cancelled

Marchete: maybe I should test that

AllYourTrees: how would you test it?

Marchete: simulating

Marchete: one single state, doing N Searches

Marchete: and see what it returns as policy

AllYourTrees: maybe the divergence between noisy probs & actual search policy?

Marchete: with low count it will be noisy

Marchete: I mean that my fear is that dirichlet for example can give diversity

Marchete: but I don't want that it was at cost of policy accuracy

Marchete: I don't know if you understand my point

struct: are you training for c4 yet?

AllYourTrees: i think i see

Marchete: not yet

Marchete: don't worry :)

Marchete: **yet**

AllYourTrees: i've only been doing c4 >.>

struct: you are still doing c4?

Marchete: with alphazero?

struct: I dont think ive updated my c4 in 27 days

Marchete: hmm robo is also too high in c4...

AllYourTrees: ya ive been mainly updating my stuff locally

Marchete: I thought it was Smits with his uber books

AllYourTrees: haven't pushed a new network in a while

Marchete: you have my code for comparison purposes :D

AllYourTrees: finally fixed a few bugs in the last week so now im getting stuff rolling

AllYourTrees: yeah :D

struct: yeah robo bot is the strongest now

Marchete: c4 seems easy to implement

Marchete: put bit -> check win

Marchete: not like the messy oware

AllYourTrees: i agree, i tried looking at oware and was like nope

struct: same

Marchete: you can win before move

Marchete: or you can lose when you think you'll win

Marchete: so I can't recommend

jacek: c4?

Marchete: Oware it's like some farmed just pulled random rules out of his a** and put on the game

Marchete: then something doesn't please him and added more artificial rules

Marchete: farmer*

Netsmile: @hyke

jacek: :thinking:

Marchete: :hugging:

Sajarin_M: hi friends. does anyone know how to change keybinding for "playing all test cases" in the editor ?

struct: ctrl + enter

Sajarin_M: can i change it

struct: dont think so

Sajarin_M: but there is an option for changing keybindings in editor settings

struct: yeah but this keybind is not related to the editor

TobiasA: if a line is parallel to the x axis and another to the y axis how does one know if they are perpendicular assuming the "person" is a computer

TobiasA: ?

BlaiseEbuth: Because the first line is parallel to the x axis and the other to the y axis ?

TobiasA: the computer doesn't like division by zero

TobiasA: how would it know then?

BlaiseEbuth: Oh yes. I always forget that's so hard to add a 'if' to handle an edge case... :(

TobiasA: what do you mean?

struct: if (0) dont divide

TobiasA: then do what?

Astrobytes: add 42

TobiasA: really?

Astrobytes: ...

TobiasA: ...

Hilekat: hello world

struct: hi

Hilekat: how are you

Hilekat: I would like to be a very good programmer

Hilekat: how can I do

struct: practice

dbdr: practice

BlaiseEbuth: ...

jacek: exercise

BlaiseEbuth: train

BlaiseEbuth: :steam_locomotive:

AllYourTrees: programmers hate this 1 simple trick to being better!

dbdr: there was supposed to be a third "practice". we failed

struct: I don't think there is any shortcut

dbdr: no Control-V you mean? ;)

AllYourTrees: didn't github release some ML thing that codes for you recently

Astrobytes: Failing a lot helps you succeed as long as you're smart about it.

BlaiseEbuth: How to enlarge your code in one week with this NN trick !

dbdr: fail smart, not hard

struct: ah yes

struct: just do a NN then you dont have to code

TobiasA: i'm still confused on the perpendicular thing

TobiasA: i can't find the answer on google

Astrobytes: and not fail soft either dbdr :P

AllYourTrees: simply construct a simple NN to easily write your small code easily

dbdr: you minimax much Astrobytes? :D

Astrobytes: A little :rofl:

BlaiseEbuth: https://www.letmegooglethat.com/?q=program+to+check+if+lines+are+pependicular TobiasA

Uljahn: ^ also you can use dot product for vectors

KiwiTae: the projection of one onto the other is a dot~

TobiasA: what about the "division by zero problem

KiwiTae: what?

KiwiTae: lol

Uljahn: eh?

**BlaiseEbuth slaps TobiasA around a bit with a large IF.

KiwiTae: you dont need to compute the slopes

FalINTOblivion0112: i got babylon tower lol

linjoehan: man this retainig water problem is a beast

FalINTOblivion0112: first legend achievement

BlaiseEbuth: I'm the only beast here.

TobiasA: then what do i need to do

nulte: grats

BlaiseEbuth: שם המפורש

FalINTOblivion0112: * FallNTOblivion slaps nulte around a bit with a large IF.

FalINTOblivion0112: how do you do that lol

Astrobytes: return of the nulte

nulte: do what?

Wontonimo: hey TobiasA, add an if statement to look for the case before it happens

nulte: im back to finish bt maybe

TobiasA: yea then what should i put in the if statement if it is true

KiwiTae: TobiasA is trolling us no ay hehe

KiwiTae: :kissing_heart:

TobiasA: i am not

BlaiseEbuth: There's no trolls here.

TobiasA: i am at my dumbest moments right now

Astrobytes: Didn't even know you had a bt bot

Wontonimo: PM me the challenge link

nulte: i dont yet

nulte: maybe by the end of today I will

KiwiTae: TobiasA you can use hough form if you worry about the /0 its not happening there

nulte: its mostly copy paste from c4 bot

Astrobytes: if (pawn at the other side) win

TobiasA: what is hough form

Wontonimo: https://www.google.com/search?q=hough+transform+formula&oq=hough+form&aqs=chrome.1.69i57j0i22i30l2.5011j0j4&sourceid=chrome&ie=UTF-8

Wontonimo: ^^ TobiasA

nulte: tobiasA step 1 to programming is learn how to search

BlaiseEbuth: Should be step 0

Astrobytes: Points deducted from Wontonimo for not providing a letmegooglethatforyou link

dbdr: step 1/0

Astrobytes: :D

BlaiseEbuth: Yeah, thx Astrobytes

Wontonimo: :D

TobiasA: i ask ppl here when i don't get the answer from searching

TobiasA: but you guys are beating around the bush

Wontonimo: :O

Astrobytes: Yes but you're supposed to keep searching with the new info, so that you learn

BlaiseEbuth: If you don't find, search more...

TobiasA: what if i never find

Astrobytes: Then it's all over.

KiwiTae: u get u , v directions from those lines just check u.v = 0

nulte: https://www.codespeedy.com/python-program-to-check-if-two-lines-are-parallel-or-not/

BlaiseEbuth: Then stop searching for the solution, and buld your own with what you have

TobiasA: sigh

KiwiTae: u cn also compute the atan2 between them

BlaiseEbuth: Or you can just copy paste the solution given by nulte.

dbdr: while True: http://chat.codingame.com/pastebin/5e916876-dbc3-4341-827c-fe9c1b1e76d7

Astrobytes: Classy!

TobiasA: thanks

dbdr: :it's_a_joke:

BlaiseEbuth: You mean the angle is involved in the 0 thing?! :o What a surprise

KiwiTae: :joy:

TobiasA: how do i rotate it -_-

**Astrobytes runs away

nulte: I dont know

BlaiseEbuth: שם המפורש

dbdr: :no_mouth:

TobiasA: don't you run a way

TobiasA: away*

KiwiTae: [[cos -sin][sin cos]]

BlaiseEbuth: Thx Kiwipedia

Astrobytes: hahaha

KiwiTae: >< im trying to pospone translating somethin in js :/

TobiasA: this is exactly what i meant by beating around the bush:confounded:

TobiasA: if only i was smart enough to know what [[cos -sin][sin cos]] means

KiwiTae: its the rotation transform

BlaiseEbuth: Stop with this bush, there's a f****n' forest behind.

TobiasA: :no_mouth:

TobiasA: i know it is the rotation transform

KiwiTae: lik ejust take a paper draw a unit circle

KiwiTae: draw the x axis

TobiasA: but do you believe that if i write that in the if statement it will work?

Astrobytes: This is rapidly becoming a derp-endicular situation

KiwiTae: and rotate the paper

KiwiTae: done

linjoehan: AAAGGGRRRHHHHH Process has timed out. This may mean that your solution is not optimized enough to handle some cases.

BlaiseEbuth: TobiasA, a serious advice: stop trying to do black magic and start thinking about a logic solution.

Astrobytes: fixit

TobiasA: i have been thinking for two days

nulte: o.o

KiwiTae: its for which puzzle?

TobiasA: i have probably gone insane

linjoehan: https://www.codingame.com/ide/puzzle/retaining-water

KiwiTae: TobiasA insane is good

Astrobytes: There is no sanity here.

linjoehan: next idea

TobiasA: insane is good except when it isn't

BlaiseEbuth: Sanity is so 20th centur'ish

Astrobytes: "Sanity, definitely my favourite sin" :P

KiwiTae: linjoehan i struggled with that one a bit goodl luck hehe

Astrobytes: isn't that one simpler than it looks at first?

Astrobytes: iirc, don't think I solved it but I remember when it was in WIP

KiwiTae: its not hard once you get it ~

TobiasA: in a grid there is a point that has two neighbours how do i know if the neighbours are in this formation >> ". :" and not this >>". . ."

Astrobytes: check to the left, to the right, and up?

KiwiTae: +1

KiwiTae: def an overkill to check for perp lines there hehe

Astrobytes: Ever so slightly lol

dbdr: train a NN

Astrobytes: A raycasting NN!

TobiasA: :pensive:

eulerscheZahl: "There is no sanity here." you surely didn't play Code of Kutulu

BlaiseEbuth: :cloud_lightning:

Astrobytes: But I surely did!

BlaiseEbuth: Of course we did. That's precisely why

eulerscheZahl: and you lost all your sanity there?

TobiasA: at this stage of insanity training a NN might not be a bad idea

Astrobytes: TobiasA: You want to check the neighbours of the points on the grid - so check the neighbours of the points on the grid.

BlaiseEbuth: That's a trap, I don't trust you

TobiasA: you did it again:tired_face:

TobiasA: the bush thing

Astrobytes: Is this for real?

TobiasA: how do you expect me to understand what you mean by that

BlaiseEbuth: שם המפורש

Astrobytes: TobiasA 06:43PM in a grid there is a point that has two neighbours how do i know if the neighbours are in this formation >> ". :" and not this >>". . ." - CHECK THE NEIGHBOURS OF THE POINT

**BlaiseEbuth give up and grabs some popcorn.

AroOmega: Just to confirm. There's not multiple CPUs on the VM that runs your code, right?

Astrobytes: Look left, look right, look up!

Astrobytes: We get only one core AroOmega

AroOmega: Gotcha. Thanks!

eulerscheZahl: you can try hyperthreading. but it's not worth the effort

eulerscheZahl: AVX can give some boost

Astrobytes: ^

nulte: performance does not matter

Astrobytes: :rofl:

dbdr: famous last words

eulerscheZahl: if your ranking doesn't matter, then you are right

KiwiTae: nulte thats what she said

Astrobytes: Kiwi turning into jacek lol

eulerscheZahl: jacek = KiwiTae?

eulerscheZahl: (╯°□°）╯︵ ┻━┻

Astrobytes: mwahaha

KiwiTae: :joy: nooo laa

BlaiseEbuth: You have to rotate your pic now kiwik

nulte: chats keep crashing for me :(

nulte: chat*

eulerscheZahl: are you using the Microsoft Internet Exploder?

Astrobytes: You shouldn't let them drive nulte

eulerscheZahl: with Adobe Crash Player plugin

Astrobytes: :D

nulte: just chromium incognito window

BlaiseEbuth: You don't want your gf to know that you came on CG?

nulte: yes, she banned me from programming :(

Astrobytes: For premature optimization?

val-6680: someone code his tron bot less than 200 lines ?:)

nulte: She thinks it might be related

linjoehan: SOLVED! but it's not the best answer I'm sure

KiwiTae: linjoehan GG

Wontonimo: val-6680 , mine is 124 lines

eulerscheZahl: solved what?

Wontonimo: only silver though

Astrobytes: Retaining Water

Wontonimo: why you ask val-6680 ?

linjoehan: my solution uses that the height is between 0 and 26. it will fail for 10^9

TobiasA: i finally did it, i just had to add "NOT" in front of my badluck

BlaiseEbuth: smart

TobiasA: the painful part is that all i literally did was add "not":cry:

KiwiTae: val-6680 i got 160lines mid gold

Wontonimo: val was asking for my src

Astrobytes: well that's 2 whole days you'll nver get back TobiasA

BlaiseEbuth: Yeah, already kicked it from #fr

Astrobytes: Shame on you val-6680

BlaiseEbuth: 🔔

TobiasA: mhm, i am a victim of time:cry:

Wontonimo: nice KiwiTae!

Wontonimo: and in python none-the-less, so one arm tied behind your back

KiwiTae: :joy:

BlaiseEbuth: 6 lines top100 legend. Can do less

KiwiTae: BlaiseEbuth you godlike golfed your solution?

Wontonimo: 6 lines? golfed?

BlaiseEbuth: No. I use C++ and I can just remove \n after every ;

KiwiTae: :joy:

Wontonimo: i'd consider a ';' a new line

Wontonimo: semantics

eulerscheZahl: for(int i = 0; i < 100; i++) {

BlaiseEbuth: Oh ? What about for(;;)

Astrobytes: lol euler

BlaiseEbuth: eZ the speedy toad

eulerscheZahl: only Fr calls me eZ

Wontonimo: Le eZ

BlaiseEbuth: And am I not?

eulerscheZahl: you are an Fr citizen how took a wrong turn and landed in world

Astrobytes: You're just BeelzEbuth. Fr language is just a cover.

BlaiseEbuth: Of course.

BlaiseEbuth: Very baguette

Astrobytes: Blaise hangs around here to improve his killcount

BlaiseEbuth: I don't kill anyone. I do business... Words matter

Astrobytes: Crushing spammers with his baguette... "BREADSHOT!"

BlaiseEbuth: -_-

BlaiseEbuth: I laughed

Astrobytes: :rofl:

Wontonimo: am i about to see a mod kick a mod ?

BlaiseEbuth: That's like the zero division

Astrobytes: lol

val-6680: 236 lines of codes :)

nulte: just use avx then you will have no problem with zero division

eulerscheZahl: https://i.redd.it/qoe0gqe9es371.jpg

Astrobytes: Which ones val? Morse? Enigma?

Astrobytes: lol

eulerscheZahl: mods are immune to kick/ban sorry to disappoint you Wontonimo

Wontonimo: nice double barrel eZ

nulte: maybe they cant kick/ban at all

BlaiseEbuth: Use roman numerals -> no problem with zero div

Astrobytes: Never divide. Simple.

Wontonimo: or only with a numerator of 1 and a denominator of a whole number

Wontonimo: i though they had a variant of fraction in the form 1/n where n>=1

Wontonimo: and it took a 3 page proof to show that 1/2 = 1/3 + 1/6

BlaiseEbuth: Maths these days...

Astrobytes: Only 3 pages eh

Wontonimo: https://en.wikipedia.org/wiki/Roman_numerals#Fractions

Wontonimo: i don't have the source for the 3 pages claim

Astrobytes: I like to leave that reading to mathematicians. I'm cool with taking their word for it mostly.

jacek: dont trust mathematicians

nulte: mathematicians created 0

Astrobytes: Don't trust the numbers either

TobiasA: i don't like 0:triumph:

Astrobytes: Pretend it's nothing.

nulte: in physiscs there is no 0

BlaiseEbuth: Right. And dividing by 0 give infinity

TobiasA: instead there are black holes

Astrobytes: You need to broaden your event horizons.

TobiasA: how?

Astrobytes: Use the 0s

TobiasA: waaaaaht

TobiasA: i'll pass

jacek: make some square root of negative number

nulte: just play paper soccer there is no 0 there

Marchete: I liked paper soccer, 5 stars

Marchete: not like damn CSB

Marchete: I can't understand why people like that

eulerscheZahl: how many stars did you give to CSB?

Marchete: 4 :rofl:

Marchete: it was long time ago

Astrobytes: fixit

Marchete: done

eulerscheZahl: i only used that star feature 1 single time

Marchete: 2 stars

eulerscheZahl: 1 star for https://www.codingame.com/training/hard/big-bang-theory---sheldons-array-puzzle

Marchete: 1 star are reserved for broken multis with ugly rules

eulerscheZahl: direct copy from hackerrank, I wanted the bot to remove it

eulerscheZahl: still rated too high to get cleaned up

Marchete: so you solved it to try to clean it up?

Astrobytes: darkhorse64: Abalone is ready?

eulerscheZahl: you don't have to solve it to rate it

eulerscheZahl: just submit 0% is enough

Marchete: ahh

Marchete: but, but you have 100%

eulerscheZahl: i solved it because the task itself is an interesting one. i just don't like direct copies

Astrobytes: Could not solve: 1 star

**Marchete looks confused

eulerscheZahl: except for vindinium :D

Astrobytes: That was at least with permission

eulerscheZahl: MIT license

eulerscheZahl: hackerrank has different terms of service

Astrobytes: Indeed

Marchete: Approved by an anonymous CodinGamer an anonymous CodinGamer an anonymous CodinGamer

Marchete: looks fine to me

Marchete: not suspicious

eulerscheZahl: these are supposed to be alt accounts of the creator

eulerscheZahl: they also approved Space Maze

Marchete: I can't believe that!

Astrobytes: :D

nulte: alt accounts o.o

eulerscheZahl: aka smurfs

Astrobytes: You don't know nulte euler?

Westicles: you need 20 ratings to delete?

eulerscheZahl: should I? wasn't around that much recently

nulte: its public class

Astrobytes: From a while ago and yes you should

eulerscheZahl: then i hereby apologize

eulerscheZahl: if nulte doesn't know me either, it would help to make me feel less guilty ;)

Astrobytes: nulte 08:05PM its public class

Astrobytes: public class

eulerscheZahl: struct?

Astrobytes: :tada:

eulerscheZahl: why?

eulerscheZahl: did he delete again?

nulte: i just come here to hide and test random stuff

Astrobytes: He accidentally gave it away some time back

Astrobytes: But an Astrobytes never forgets

eulerscheZahl: wait, struct has a girlfriend? when did that happen?

nulte: nah it was joke

Astrobytes: I'm not familiar with his personal relationships tbh :D

Astrobytes: I'm still proud of the premature optimization joke.

Westicles: 2 days on top of golf now. I thought they would take me out by now

Astrobytes: well played Westicles. That effort has payed off it seems. Rather you than me tho lol, I'm not masochistic enough to put myself through all that :D

darkhorse64: Yep, Abalone looks ready. I have not yet written a bot for it. My submit is a slightly modified copy of the boss

nulte: your solution is in bash=

nulte: *?

Astrobytes: darkhorse64: ah ok, I've not written anything for it yet

Westicles: I think pardo is the only one who cares enough to whack my scores even if it doesn't improve his, so once he notices I'll fall back

Astrobytes: Kudos still, a laborious task.

darkhorse64: What a jump ! You found new tricks ?

KiwiTae: gg Westicles

Westicles: Not really, mostly just optimizing ones from the easy puzzles

Astrobytes: Deep diving into the netherworld of each language is not always fun

BugKiller_328: Hi

BugKiller_328: Can I ask basic question about constructor ?

BugKiller_328: http://chat.codingame.com/pastebin/07770a80-0e88-4d8c-88cf-00667809f4e9

BugKiller_328: what's the difference between 1st and 2nd in making constructor

BugKiller_328: I saw that it worked differently so I'm asking to know.

BugKiller_328: 1st. that default constructor is called even I call another constructor with parameter.

BugKiller_328: 2nd is not called.

BugKiller_328: sorry nevermind..

BugKiller_328: it seems like same for 1st way and 2nd way to make default constructor.

Wontonimo: hey BugKiller_328

Astrobytes: What language?

BugKiller_328: c++

BugKiller_328: sorry, I think it's same.

BugKiller_328: yes Wontonimo!

Astrobytes: in future, post real code. Otherwise we'll debug the lack of semicolons or something.

Astrobytes: Also, there are no access specifiers

BugKiller_328: Astrobytes yes I'll.. thanks.

Astrobytes: :+1:

nulte: im not sure there is any diference

Astrobytes: other than a compiler warning from VS then no

BugKiller_328: yeah, nevermind. it's same. I've just confused myself.

nulte: I use the 2nd way though

nulte: it allows me to do

nulte: A(int data):data(data){}

Astrobytes: same, it's cleaner

BugKiller_328: yes, I always used 2nd, but today I saw 1st way and I thought it could another functionality at first.

BugKiller_328: :)

Astrobytes: and that works, sadly. Which is confusing.

Astrobytes: @nulte not @BugKiller_328

Astrobytes: Sometimes I just think RIIR is a good idea, then I try to RIIR

nulte: whats RIIR?

Astrobytes: Rewrite It In Rust

Astrobytes: not to be confused with RAII

nulte: whats RAII ...?

Astrobytes: Don't test my patience :P

jacek:

Astrobytes: do you have an abalone bot jacek, a real one that is?

jacek: no

Astrobytes: Meh. Seems ready for approval. Don't do it.

jacek: hm?

Astrobytes: Don't just blindly approve it I mean.

CSharperDude: Hi there, I just signed up

CSharperDude: im testing my c# skills but I cant find a way to test my program to get the debugs, any help?

Astrobytes: print to stderr

Astrobytes: Or Console.Error in C#

CSharperDude: but i dont see a run button, to actuallrun the code

CSharperDude: only the test cases buttons

CSharperDude: which only get red since is not complete yet

Astrobytes: That runs your code

Astrobytes: Check this: https://virtual-atom.com/codingame/introduction/

Astrobytes: Use Console.Error to output debug info when you run a testcase

CSharperDude: thank you,

Astrobytes: You're welcome

Kozakx: i dont understand what to do

Astrobytes: Who does

Kozakx: no, seriously, i move my pod and before it reaches 1st checkpoint it says it's too slow

Astrobytes: Go faster?

Kozakx: how? i set thrust to 100

Astrobytes: Maybe you're aiming slightly wrong

Kozakx: how do i get coordinates of checkpoint?

Astrobytes: you get the coordinates of the next checkpoint, you can store them

Kozakx: but where are the coordinates?

Astrobytes: checkpointX, checkpointY or something similar

Astrobytes: it's given in the input. But only for the next checkpoint, if you want them all you have to save them

Astrobytes: *nextCheckpointX, nextCheckpointY, something like that

Kozakx: ok i see

Kozakx: but i cannot get through that, just says im too slow

Kozakx: it starts the code and says im too slow

Astrobytes: paste your code, sounds like you might not be outputting anything

Kozakx: i just have the first command

Kozakx: print(str(next_checkpoint_x) + "100 " + str(next_checkpoint_y) + " 100")

nulte: print(str(next_checkpoint_x) + " " + str(next_checkpoint_y) + " 100")

Astrobytes: ^

KiwiTae: you are missing spaces

Astrobytes: and printing 100 twice

KiwiTae: print(next_checkpoint_x, next_checkpoint_y, 100)

Kozakx: it works now, thank you

Astrobytes: :tada:

nulte: bt taking a while to do

Astrobytes: I'll happily take any optimisation tips for that

nulte: ill see what I can come up with

nulte: I was doing a mistake of getting all the moves on rollout

nulte: which is a bit expensive

Astrobytes: I still have alphabeta, I go between 28th and 21st

nulte: Seems good

Astrobytes: There are other things I can do but I was thinking MCTS ept might just be the easiest

nulte: I never did ept

Astrobytes: Once I get a good alphabeta going, switching to EPT isn't so difficult

doomento: Guys, I think I will master python

doomento: and I figured out why I can't solve these problems

doomento: I wasn't thinking programatically

Astrobytes: Don't think. Feel! It's like a finger pointing at the moon.

doomento: I still don't know how to get [highest]

doomento: cause I can't think

doomento: :astonished:

KiwiTae: whats the max "highest" between 1 and 10?

doomento: 10

nulte: seems like a trick question

nulte: how did you get 10?

doomento: because they said max so I thought get the highest number between 1 and 10

nulte: do you know how to do that in code?

doomento: I can uh, try

nulte: checking 2 numbers is quite simple in most programming languages

nulte: if (a > b)

Astrobytes: result = max(a, b)

nulte: yeah that is simpler

KiwiTae: result = (a+b+abs(b-a))*.5

nulte: :nauseated_face:

KiwiTae: -

ANONYMOUS42: print(10)

doomento: :fearful:

Astrobytes: solved it yet demento?

doomento: no, I'm still trying a bunch of other things

Astrobytes: everything other than getting the max right :smirk:

doomento: I'm supposed to do that?

doomento: nvm

nulte: well you did say you couldnt get the highest

doomento: ok ok:spy_tone4:

Astrobytes: maximum and minimum, get those concepts down

Astrobytes: It's when you have a huge mother and a tiny mother

nulte: check the highest bit that is not present in the other number

nulte: but care on negative numbers

Astrobytes: struct

Astrobytes: nulte I mean

nulte: astro

Astrobytes: Is there a faster way to do this: _tzcnt_u64(_pdep_u64(1ULL << (rnd64.getRandomInt(_mm_popcnt_u64(board))), board));

nulte: whats _tzcnt_u64?

Astrobytes: ctzll

nulte: is that to get a random move?

doomento: So I have to get the highest mountain and then shoot it. Since mountain_h is the mountain height I need to find the index of the highest mountain then shoot it right?

Astrobytes: aye

nulte: usually i dont need ctzll

nulte: because im getting the mask already

nulte: then I can simply do board |= move

nulte: does that make sence?

nulte: currently a move for me in breakthrough is just 2 set bits on a uint64_t

Astrobytes: Yes. I need to re-read a few

documents tomorrow

nulte: then I can do gs.boards[current_player] ^= move; gs.boards[current_player ^ 1] &= ~move;

nulte: But im not sure how well this does

nulte: I havent finished the bot yet

Astrobytes: that makes sense. I currently have mine as 2 uint8 (from, to) and a score (helps with ordering)

nulte: yeah it makes sence for minimax

nulte: sense or sence?

Astrobytes: sense

nulte: thanks

Astrobytes: np

Astrobytes: and for some reason I use bextr to get stuff auto from = _bextr_u32(mv, 0, 6);

nulte: also I dont think there is potential for avx in bt

Astrobytes: not for anything useful, eval maybe?

nulte: Im not sure its easy to eval moves ehre

nulte: here*

nulte: but I could be wrong

Astrobytes: rank placement of the pawns is important

nulte: isnt it more important to keep pieces together?

Astrobytes: yes there's the protection part, but think in terms of attack and defense, where is the threat?

nulte: Do you take into account other pieces placement when evaling mvoes?

Astrobytes: Yes

nulte: ah makes more sense then

nulte: I thought you were evaling based on piece position only

Astrobytes: I should work a little more on the eval tbh.

nulte: I wonder if quiescence search can help here

Astrobytes: Possibly. I had a little improvement with using killer move, I tried non-capturing only but that doesn't work here (not for me at least)

nulte: what is killer move?

Astrobytes: easier to link to it than explain: https://www.chessprogramming.org/Killer_Heuristic

nulte: I see

Astrobytes: Sorry, I'm quite into my alphabeta minimaxes :)

nulte: its totally fine

nulte: I like to do simple mcts

nulte: and then optimize and add a few things to the rollout to make it smart

nulte: my c4 is the example of it :D

Astrobytes: You may as well do EPT in some cases. Where you at now in C4? Still chasing robo?

nulte: nah, was never a fight tbh

nulte: I havent improved it

Astrobytes: Yeah, looked at the Oware board recently? :(

nulte: yeah but its a good thing

nulte: One day ill try to do a NN aswell

Astrobytes: Yeah same, it's really cool to see actually

nulte: only way I could fight him is by adding a book

nulte: but even msmits is struggling with it

Astrobytes: lol, then you know it's trouble

nulte: There is one more person with book

nulte: but is only 3 plys or something

Astrobytes: I didn't look recently, who has it?

nulte: fica

Astrobytes: ah right

Astrobytes: I remember from the chat actually

Astrobytes: Oh yeah, my _tzcnt_u64 thing is from Clobber, not BT

Astrobytes: Still have work to do on that one

nulte: I still use i though on c4

nulte: but only on expand

nulte: So I dont need to store the mask on the node

nulte: 1 byte vs 8 bytes

Astrobytes: Yeah, I need to work on my clobber nodes tbh

Astrobytes: Refactoring needed

Astrobytes: Anyway, I gotta shoot. Take care man, see ya tomorrow

nulte: gn

Raibatsu: Chuck Norris puzzle has exactly 80,000 solvers. idk if I should solve it and ruin that perfect number though :(

MorningStar37: You must, because you can.

xiaoyaozi: go run main.go