Chat:World/2022-07-10

From CG community
Revision as of 05:29, 11 July 2022 by Chat Log (talk | contribs) (Created page with "<img src=/a/82297665972597> gokubill: hi <img src=/a/75369242083208> jacek: oO <img src=/a/82297665972597> gokubill: spam spam =]] <img src=/a/16542412157988> Zorg1: nice...")
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

gokubill: hi

jacek: oO

gokubill: spam spam =]]

Zorg1: nice gn spam ;-)

aCat: morning

aCat: (late morning I mean ;p)

j4at: morning

jacek: j4at overnight training brought me to 6th :tada:

j4at: :tada:

eulerscheZahl: for me the night brought me to 10th already

Default avatar.png putibuzu: gz

jacek: bz2

MSmits: lol, this submti

MSmits: submit

MSmits: http://cgstats.magusgeek.com/app/multi-tic-tac-toe/msmits

MSmits: if i spam this bot, jacek is 2nd and robo goes out of the top 10

jacek: spam plx

MSmits: hehe sorry, just experimenting. I think this is a good one, but it's hard to say when most of its rating comes from bashing robo

jacek: good enough

MSmits: I've had one before that did the same to you. I wonder if that one was better

jacek: im at the about its 100% win against my test bots https://i.imgur.com/xXsANfw.png

jacek: its improving but i cant see it anymore

MSmits: looks good

jacek: at the point*

MSmits: make more testbots

jacek: i think i can get to top3 with those inputs. i thought using inputs for the whole board would be required

MSmits: i can't comment on that, i havent tried to understand the game yet

jacek: i dont understand that game either ~

MSmits: right, but if i did, i could help you maybe. Too busy with crosses and circles though

Default avatar.png IonutH: hi. what happens if I submit my code with 0% result in Clash of Code? How does affect my rank? What if others do the same in a game? Do I have an advantage if I do it fist, meaning to give up?

jacek: its based on rank. if youre 1st with 0% then you win

jacek: or if shortest, 0% with 0 bytes wins i think

Thorcode: 0 bytes in shortest lose

jacek: oh

Default avatar.png IonutH: thank you for answers

Thorcode: you must type a letter

Thorcode: if you give up

MSmits: an apology letter?

Thorcode: in shortest if you give you just delete all and type an letter

jacek: thats short for sure

Default avatar.png IonutH: oh nice hint

MSmits: err, it's better just to not lose :)

Thorcode: it give you a chance to win

MSmits: I guess

j4at: jacek what hardware are you using for training?

MSmits: he's using a chromebook :P

j4at: yeah, probably :p

jacek: cpu ryzen 7 5800X. i have 2060 gpu but i didnt managed to make it work to my needs yet

j4at: my dude have a gpu and using the cpu instead :p

jacek: :shrug:

DomiKo: j4cat gpu is useless

DomiKo: trust me

j4at: no its not

DomiKo: yes it is

j4at: it speeds training a lot

DomiKo: not really

j4at: huh ????

DomiKo: its speeds when you network is big

jacek: maybe for big deep nets

j4at: use a big batch size

j4at: problem fixed :)

DomiKo: but for small networks cost of transfering memory from CPU to GPU is much higher then rest

DomiKo: no

DomiKo: it's not about batch size

j4at: you do'nt need to transfer it tho

jacek: + big batch size leads to poorer training

DomiKo: true

Default avatar.png RafIntalian: Hello everyone

j4at: well it really depends

MSmits: hello RafIntalian

j4at: for exemple openai is using 2million batchsize for gpt3 :p

j4at: There is some tricks to make larger batchsize better

jacek: jax or evojax looks promising for many small nets

MSmits: jax? shortened from jacek?

j4at: I'm using 512 batchsize for my search race

MSmits: jaxmax

DomiKo: 512 is not much, but when you have many millions of samples, you can't fit them into GPU, so you have to transfer

j4at: heh ? you said small neural networks

DomiKo: network is small, but you have many samples

j4at: That is not costly really

j4at: I keep all samples in the CPU

DomiKo: transferring the memory?

j4at: transferring samples yes

j4at: because we are not dealing with images here

DomiKo: images?

DomiKo: for me memory transfer is around 80% of the whole GPU time

j4at: maybe you are doing it wrong

jacek: plug your gpu closer to ram

DomiKo: PyTorch lighting is doing everything for me

DomiKo: so I can't really do something wrong

MSmits: lighting? I thought it was for training NN

j4at: maybe your gpu is connected to psu x1 or something

j4at: lmao

DomiKo: MSmits yeah lighting

MSmits: I guess usually training NN gives heat, but I guess pytorch is different

MSmits: is it natural light? Or more like LED?

jacek: its optical computer

DomiKo: xD

DomiKo: it's lighting fast!

MSmits: I am starting to get the sense you mean lightning?

DomiKo: ohh yeah

j4at: did a test. in gpu it takes .05s for every "iteration" (training 4 nns) and in cpu it takes 1s even if i made it train 1/2 nns it will still be faster

j4at: Maybe you have gtx 1030

j4at: or using an igpu

j4at: with 2 nn it takes .025s proving that data transfer almost takes no time in this case

Default avatar.png putibuzu: how large are the NNs?

j4at: Well its hard to explain. I will make them even smaller

Default avatar.png putibuzu: what data is being transferred here? training samples? and the weights are already stored in GPU memory?

j4at: yes

Default avatar.png putibuzu: ok thanks :) I was always wondering how people make use of GPU for training

Default avatar.png putibuzu: to clarify do you use the GPU for forward pass as well?

j4at: i use gpu foreverything yes :)

jacek: to make coffee as well?

j4at: yeah i use the extra heat :p

Default avatar.png putibuzu: ok then you basically only have to transfer the weights at the beginning of training and when writing to disk right? and the weights are modified during training in the GPU memory

MSmits: and can it do lighting like DomiKo's ?

j4at: putibuzu The weights gets initialized directly in the GPU

Default avatar.png putibuzu: j4at alright thanks!

j4at: They are critic nns so they don't need to fit in cg because i'm using the actors. But even when i made them small enough to fit in cg. gpu takes .01s and cpu .06s

MSmits: will it use the same amount of energy in total, to train the NN?

j4at: less energy

MSmits: hmm

j4at: because gpu is more energy efficient

jacek: if you managed to work with your gpu, fine. i didnt

j4at: and when I train it on small nns like this it uses like 10% of its power

j4at: With pytorch/keras its pretty easy

jacek: https://i.kym-cdn.com/photos/images/newsfeed/001/450/824/6a8.jpg

MSmits: is that shaggy?

j4at: I didn't use mixed-precision either with mixed-precision you can get 2x the speed

jacek: MSmits yes

MSmits: 495 submits in uttt

jacek: ban :hammer:

MSmits: (over 4 yrs)

DomiKo: j4cat

DomiKo: https://colab.research.google.com/drive/1y7a0_ovqgtNDikXKr4rapEqJFlQ1Ksde?usp=sharing

DomiKo: simple example where cpu = gpu

j4at: colab gpus are trash

j4at: they are randomly choosen

DomiKo: cpu are too

j4at: you can get a good gpu or a very slow one

j4at: do the same test in kaggle

j4at: and send me the results :p

DomiKo: even in simple test when I move whole data set to cuda and then train, it takes 1s less (from 8s)

DomiKo: so yeah that proves my point that transferring that is very costly

DomiKo: and I never used kaggle, so I can't do it

j4at: That model is less than 40k weights half my nns size. I will test it on my computer and send results

DomiKo: it was just a simple test

j4at: well, I timed only the training not moving the model too. And the results are 4.96s in gpu and 5.87s in cpu. 512 inputs is more than my 112 input and my model is 2 times bigger.

j4at: also you are using 1 core to load the data

j4at: I will try using all cores

j4at: 3.9s in gpu and 4.45s in cpu

DomiKo: still pretty close and have have new GPU

DomiKo: " 512 inputs is more than my 112 input and my model is 2 times bigger." as I sad it's just a simple example

DomiKo: I just typed some random numbers

DomiKo: but especially with one hot inputs, you can have huge input layer compared to the model size

j4at: Did some optimizations and now its 3.78s, yeah maybe you are kinda correct DomiKo gpu is only 2 times faster than cpu in this case even after I increasesd the size to 80k weights

j4at: without changing input size

DomiKo: and with CPU you can train in parallel

j4at: IN gpu too

j4at: atleast using pytorch

j4at: but for actor critic gpu is a lot faster. Because you have 2 nns + you can make the critic as big as you want

jacek: youre so critical

j4at: + if you use attention or cnns other other fancy operations cpu takes a lot a lot longer

DomiKo: yes it would

j4at: "have new GPU" Domiko I have gtx 1080 so I don't have tensor core :'( I rent gpu's online :'(

DomiKo: :cry:

MSmits: http://cgstats.magusgeek.com/app/multi-tic-tac-toe/msmits This version is a NN-buster. only robo unscathed

MSmits: it's funny that when you have enough different versions, you can just decide who is nr 1-2-3

jacek: one man shall not have that much power

MSmits: :grin:

MSmits: could have kicked recurse out of top 10 with that too

DomiKo: LOL

MSmits: 17 W / 6 L

MSmits: 1 D

MSmits: if I had more random in my bot this would not have been an option. The fixed eval and no random rollouts really makes this happen

MSmits: i should keep a library of such bots for each arena and then when someone annoys me, I'll submit my destroyer bot for that specific person :P

MSmits: it's evil

j4at: jokes on you my bots are not good enough to reach the top.

Default avatar.png putibuzu: :D

MSmits: lol

DomiKo: MSmits so you will be like an ultimate mod?

Default avatar.png Aldoggen: the Great Mod

Default avatar.png Aldoggen: aka God

MSmits: sure why not :)

MSmits: it's weird, because my bot is not strong enough to get nr 1. 4 at most, if i am lucky.

j4at: Can i join your cult ?

MSmits: sure

j4at: :)

DomiKo: MSmits 4th is still impresive

DomiKo: and it really shows how strong karliso is

MSmits: Yeah

MSmits: personally I think he is the best bot on the board. I can make versions that beat any NN, but i cant do the same for karliso. His bot is more random

jacek: should i put more random

jacek: :imp:

MSmits: you could, but it might make your bot weaker

jacek: as long as its anti-MSmits

MSmits: mmh it wont help that much, when i wreck robo or recurse, it is by a lot. Yours already does reasonably well

MSmits: because you already have the random

MSmits: my bot is also just around your bots rating, similar strength. without any RPS

MSmits: I think yours is a little bit stronger

MSmits: oh and wait till you start experimenting with anti-msmits stuff. I should be done fitting params after today. I'll be generating a book after that, but that takes weeks

struct: NNs are over rated I knew it

jacek: referee should tell you opponent's name. that would make things easier

MSmits: so true

MSmits: it would be funny if your bot could read opponent msg output

MSmits: identify opponent that way

MSmits: well, there sometimes is a way, if you want to go that far. In many games, the top bots have a specific but different first few moves. You could detect them that way

MSmits: I never did that personally

MSmits: doesnt work on onitama and such of course

j4at: download top 20 players games and train a NN to identify the player

MSmits: lol

MSmits: thats some meta machine learning

jacek: https://maiachess.com/

StevenV: hello guys

MSmits: jacek now make a chess engine that plays like a monkey

struct: Maybe they can play chess

MSmits: they can play some games

MSmits: I don't think banana-motivation is enough to get it to learn the rules of the game

StevenV: I am gonna learn kotlin, is it good for COC?

eulerscheZahl: no. but that's a strange priority to learn or not to learn a langauge anyways

struct: I dont think so

struct: MSmits have you ever seen them doing chimp test?

jacek: kotlin compiler is slow here and might even timeout

jacek: https://en.wikipedia.org/wiki/Infinite_monkey_theorem#Actual_monkeys

j4at: StevenV try vb.net jacek recommended it to me.

StevenV: I am an experienced Java dev, that's why I want to learn Kotlin...

MSmits: i saw some stuff, not sure if i saw that struct

j4at: hehe :)

jacek: try fortran or cobol. much future

StevenV: :(

jacek: kotlin is fine. its just weird for clashes

DomiKo: for clashes everything should be fine, except shortest...

StevenV: thanks guys, I guess I just use kotlin to solve puzzles then

j4at: yeah only shortest... fastest and reversed :)

Thorcode: using clash to learn new language

Thorcode: clash is good to learn new language

MSmits: made it to 3 :)

MSmits: http://cgstats.magusgeek.com/app/multi-tic-tac-toe/msmits

MSmits: all NN's matched or busted :)

Thorcode: cool

MSmits: damn that karliso though =)

Thorcode: just made it to bronze

Thorcode: that tic tac toe game is hard

Thorcode: due to changing board

MSmits: yes, its a huge change

MSmits: i would have preferred some leagues where the normal TTT boss just becomes better

Thorcode: through wood ez

Thorcode: but bronze like a lot of pain

MSmits: the simulation is still quite simple though

MSmits: it's just hard to know what is a good move

Thorcode: sometime

MSmits: i have been toying with the idea of making a new UTTT contribution with better TTT leagues and final league has steal as an option

Thorcode: I could get to bronze by brute force

MSmits: then you dont need mirror matches either

Thorcode: oh that's seem fun

MSmits: I never made a game before though, on CG

MSmits: and have to find the time

miszu: Hey all, are there any competition where it's all based on heuristics and none on wiriting bulletproof simulation and engine?

MSmits: you mean bot arena's?

MSmits: yes plenty

miszu: yeah

MSmits: code a la mode is closest to an all-heuristics competition

miszu: I find it teadious writing an engine.

miszu: I just want to read the inputs and do the best heuristic

MSmits: as for tedious, my code a la mode bot had 3k lines :P

miszu: bruuh

Thorcode: how

MSmits: i did have search but, so many ifs

MSmits: over 300+

MSmits: you did ask for heuristic arena's

miszu: I know

MSmits: means many ifs

MSmits: = much code

miszu: as long as I don't need to sims

MSmits: the winner didnt have sim

MSmits: of the contest that is

MSmits: python winner

MSmits: I ended 8th with beamsearch

miszu: so I can get to play with just heuristic?

MSmits: i was an exception, most did not use search

miszu: plat*

miszu: I mean legendary lol

MSmits: easily

MSmits: well

MSmits: write good heuristics :0

miszu: I will :0

MSmits: and pick up my damn plate

MSmits: :rage:

Thorcode: brute force

miszu: Don't tell me what to do

miszu: I am an unionized coder

miszu: we have rights

MSmits: miszu in the contest i would put down finished plates, that had to be carried to the dropoff and when the other bot was closer, i would attempt to pass it to them

MSmits: but i got ignored a lot

MSmits: (it's a cooperative arena)

miszu: Interesting

miszu: I will have a try then

MSmits: it's really fun imho

Thorcode: gotta learn something new

Thorcode: bye

MSmits: bye Thorcode

miszu: I was doing coding royale but when I saw to go to legendary you need to write a buggy sim and engine...

miszu: I was like life is too short for that

MSmits: heh i can get that

MSmits: sims for simple boardgames are not a lot of work though

miszu: yeah that's fine

miszu: but like the pod racing?

MSmits: yeah thats work

MSmits: and there are worse ones

miszu: the fun part is the heuristics

MSmits: sometimes it's fun

MSmits: sometimes i just dont know what to use

MSmits: last contest i didnt know

MSmits: i just bruteforced my own turn and got to legend that way

miszu: I wish we could do a manual mode where you just play the game yourself

MSmits: yeah that would be fun

MSmits: jacek, good news, you rank above recurse and are still 4th

Default avatar.png HenroLST: morning

MSmits: I'm done fitting params, so if you want, mess around to create anti-smits, but you'll push recurse back up, that's the price

MSmits: morning HenroLST

Default avatar.png HenroLST: geez how are you guys always awake and coding

Default avatar.png HenroLST: are you like lucid dreaming just to code in your sleep too???

MSmits: only during contests

MSmits: this is me relax-coding

Default avatar.png HenroLST: dang

Default avatar.png HenroLST: ok i need to up my game then :sweat_smile:

MSmits: just enjoy and you'll code more

Default avatar.png HenroLST: yup

Default avatar.png HenroLST: wait just out of curiosity can you make a NN in c++?

MSmits: you can make a NN in anything

Default avatar.png HenroLST: ooo something to look forward to in the future then

MSmits: generally it is easier to use one of those python frameworks you can find

MSmits: you can also export the weights you get with those and then use them in a c++ bot

MSmits: thats what many do on CG

Default avatar.png HenroLST: ah

Default avatar.png HenroLST: or just grab from stack overflow :stuck_out_tongue_closed_eyes:

MSmits: but in the end a NN is just a program, so it's not language specific

Illedan: During offline training. Is it better to test against a constant population or create a new population to do round robin against each round?

MSmits: for testing probably better to have constant set of bots

MSmits: otherwise how will you know you improved

Illedan: Yeah, I'm getting stuck on that thought now :thinking:

Illedan: So I need to figure out how to make a good variety of bots :D

jacek: that was my problem in green circle

jacek: later i used bots trained with different hyperparams for testing

jacek: at the beginning i was testing against random and the 1st action bot only

jacek: maybe if i knew the game i'd make some simple heuristic bot as well

Illedan: Now I'm creating a new population of 4 new mutated bots + the best from the prev round. Then doing 20 games between all pairs (80 games for each)

Default avatar.png D4WEED: (╯°□°)╯︵ ┻━┻

Illedan: This is for FB

jacek: NN? :scream:

Illedan: Nah, GA hyper param training

Illedan: Could put them through some NN though

jacek: im doing pure round robin, coevolution for clash of bots. i just ran out of testing bots, its 100% wins against all of them now

jacek: https://i.imgur.com/VBecERg.png

Default avatar.png HenroLST: bot

Illedan: I'll change to a population of 2 and just save all the new best mutations. To get a validation population

jacek: sounds incest

Illedan: My newly trained IAs have dropped re Curse 2 points since yesterday. So soon I can have hopes of being first :eyes:

jacek: so incest is wincest?

Illedan: Only the good kind of incest

Default avatar.png HenroLST: ...

MSmits: jeez, I am away from the computer for 1 second...

Illedan: Hm, I'm talking about my GA genomes

MSmits: o

jacek: and youre inbreeding them

Default avatar.png HenroLST: imagine reCurse comes back just to win that particular event

struct: he is retired

Illedan: Did he say?

Illedan: Or just asumptions?

struct: it says on his CG profile

struct: retired from CG

Astrobytes: and he did say he was done

Illedan: I guess his last contest win was enough to prove everything

struct: I guess I'll never retire

jacek: he said not-first in contest or retire

karliso: MSmits wow, look at you, what have you done to uttt leaderboard?

Astrobytes: probably concentrating on chess, and little recurs e too of course

struct: karliso his bot is now eval instead of rollout

karliso: yeah, I wonder how hard it is to come up with good eval.. or just that noone gave it serious try.

Illedan: 1 random move and then eval all the way?

struct: eval on expansion Illedan

Illedan: Which is?

struct: MCTS select expand on expansion when you create a node you eval it

struct: instead of doing a rollout

Illedan: And then do rollouts afterwards?

struct: yeah no rollout

struct: just eval straight away

MSmits: karliso yeah i messed it up :)

struct: Am I right MSmits?

Illedan: I don't quite understand

MSmits: yes struct

MSmits: just dont do the random rollout in mcts

MSmits: replace it with eval score

MSmits: and backpropagate that also

Illedan: Ok, so you the rollout to some depth?

Illedan: And then eval instead?

Illedan: *you do the

MSmits: my depth for this is 0

MSmits: i expand, then eval immediately

Illedan: So it is instant?

MSmits: yes

MSmits: selection -> expansion -> eval

Illedan: Done in 1 ms?

MSmits: eh no i do this 100k times per turn :P

struct: he does normal mcts just no rollout

Illedan: hmm

MSmits: to be clear, 142k on 2nd turn

jacek: jacekmax or mcts?

MSmits: eval is pretty expensive

MSmits: not jackemax no

MSmits: I dont know how that would do

Illedan: Got any documentation that I can read?

MSmits: nope, but it really isnt complicated. Imagine a normal mcts

DomiKo: MSmits with eval you have more itererations?

MSmits: at some point you have a line that says score = Simulation ()

MSmits: replace that with

MSmits: score = Evaluation(0

jacek: mcts early playout termination, though they play for N random moves, then eval instead of checking win/loss

MSmits: DomiKo a little more yes

jacek: more in the beginning but probably less in ending

MSmits: but the eval contains a lot of ifs and lookups so its not a lot

karliso: So you normalize eval to -1 1 or equivalent?

MSmits: i dont normalize

MSmits: what i did was, i keep the exploration param at 1.0

MSmits: and fit the other parameters to something that works

jacek: would switching to rollouts from late midgame be better?

MSmits: it could totally be the case that there are regions in the game where rollout is better or eval is better

MSmits: karliso the working params are such that the scores more or less fall into the same range. Sometimes they go at most to -2 or +3 or something

MSmits: btw, it was not hard to make an eval that gets into top 100 for uttt. a simple eval that makes sense will get you there. A good eval is harder

karliso: How long did it take for you?

MSmits: i had this bot from a year ago, i worked on it then a week or so, but i gave up to soon. I picked it up now and improved it in a few days

MSmits: the other bot was ranked around 15th

MSmits: (the earlier eval bot i mean, mcts was ranked 8)

MSmits: 8 lol... damn emotes

MSmits: but just the fact that a neural network can be trained to eval uttt should tell you it's possible.

karliso: But beating NN is a different thing.

MSmits: i am not sure it is an entirely fair comparison. I suspect if you look at the games, they are often the same game

MSmits: if i change 1 param by 0.1, suddenly recurse wrecks me, or robo does

MSmits: if both bots do not use random, it can get kind of boring

MSmits: but the bot does well vs other bots too, so i wont complain

jacek: :rage:

jacek: im so happy for you

MSmits: I can tell :P

MSmits: dont worry, i am sure i'll get pulled back down when the other top 10 bots resubmit.

MSmits: the nn's tend to do better vs random bots

MSmits: vingte, tomatoes or yurko can ruin my day

MSmits: http://cgstats.magusgeek.com/app/multi-tic-tac-toe/msmits

MSmits: i will not get much better than 50% and the NN's will much more

_NikJ: Now, what I'M wondering is, what kind of magic did reCurse cast in order to get such an insanely high score in mad pod racing?His win ratio is like 96% lol

jacek: we may never know

MSmits: _NikJ mad pod racing is kind of strange in that sense though. If 1 bot is better than another, even if only a little, they quickly get 100% wr

MSmits: only a small margin in bot strength is necessary to get very high WR

MSmits: the same margin in uttt would get you maybe 55%

Default avatar.png putibuzu: MSmits what techniques do you use for parameter fitting?

MSmits: crap techniques

MSmits: i tried first locally playing vs my normal mcts bot

MSmits: i figured it wouldn't overfit but it eventually did. It would even wreck my mcts both with 500 ms calc time

MSmits: so then i just tried stuff on the leaderboard

MSmits: (submit)

MSmits: thankfully I only have a few params now

MSmits: I think 3 main ones, the other ones are pretty basic and can be independently fitted (similar to exploration param)

MSmits: for example, I have a mobility param, that subtracts from the score. childCount * param

MSmits: that doesnt really depend on anything else I think, so i just fitted it once

MSmits: (it's bad to give your opponent more choices)

Default avatar.png putibuzu: ah ok thanks :)

MSmits: if i had to do it better, i would probably do a GA. I should make a frame work for it

MSmits: it would be nice to have one for next contest

MSmits: where you just plug in your bot with the ability to pass param value and then the framework starts them fighting

Default avatar.png putibuzu: hat sounds interesting :thinking:

I was wondering whether you can use self-play with stochastic gradient descent like alpha zero

Default avatar.png putibuzu: *that

MSmits: yeah i think there are some ways like that, but GA is fairly simple and works well. I just haven't done it yet

Default avatar.png putibuzu: Alright thanks :) I'm looking forward to you taking #1 :)

MSmits: not by fiddling more params, but i am generating an opening book. That should help

MSmits: I am not counterbooking though

jacek: books :unamused:

MSmits: I am not feeling guilty for using them when others spend days locally training their NN and filling their codesize with weights :)

MSmits: counterbook is cheesy though

TimeEngineer: :joy:

Illedan: I wonder if I can get into UTTT legend by only using books?

MSmits: books are really weak in uttt, there are better games to use them

MSmits: problem is that there are many ways to play well, so there's many plays to counter

MSmits: game winning/losing mistakes happen somewhere between ply 18 and 30 for good bots

MSmits: and books dont usually go that deep

MSmits: I am just hoping for a few % WR from them

Illedan: How many moves can you fit in a book?

MSmits: depends on the game

Illedan: UTTT ofc

Illedan: I know it depends :P

jacek: not many transpositions there

MSmits: i am thinking

jacek: :scream:

MSmits: probably around 200k

MSmits: but you'd need to minify your code to a negligible amount or it'll be less

MSmits: i have never hit the limit though

Illedan: Maybe I'll try for UTTT

Illedan: See if pure lookup can go to Legend :D

Illedan: Can't do anything while FB is training locally

MSmits: well, you wont be able to do pure lookup. Still need some heuristics to finish off

Illedan: Nah ^^

MSmits: you wont be able to finish a single game just with lookups :P

TimeEngineer: What's difference between Qtable and books ?

MSmits: it's the same thing i guess

TimeEngineer: I see

MSmits: Qtable stores states with actions right?

TimeEngineer: Yes

MSmits: yeah same thing

MSmits: I dont store the states though

MSmits: just the moves, in a tree starting from the root

MSmits: moves are much smaller than the whole game state

jacek: how do you deal with transpositions

TimeEngineer: That's the question

MSmits: I deal with it locally.

MSmits: I don't generate moves in the tree that are duplicate

TimeEngineer: Many actions can lead to the same state

MSmits: if it hits a gamestate i find in another branch, then it ends there, as a leaf

MSmits: locally when generating book, i take into account rotations, but also other kinds of transpositions

MSmits: when in my bot, i need to lookup a move, I try all rotations of the current state, to see if it matches with a state from the book

TimeEngineer: Do you use NN for UTTT jacek ? if so you used a CNN ? I'm wondering because of rotation invariance

jacek: i use NN, but good old one-hots

jacek: i just add rotations to training positions

MSmits: did you do the inside-out transposition for nmm?

jacek: the what

MSmits: switch the inner ring with the outer ring

MSmits: it has 16 ways to transform the board

TimeEngineer: What XD ?

MSmits: nine mens morris

jacek: no, just rotations and symmetries

TimeEngineer: Ah it's another game :p

MSmits: ahh thats sad jacek, could have doubled your training states =)

jacek: i have no motivation, im still 1st :(

MSmits: true

MSmits: TimeEngineer uttt has 8 symmetries, yavalath has 12 and nmm has 16, it is kind of special that way

MSmits: besides it being bitboardable with only 24 bits per player

TimeEngineer: I have never played others game

MSmits: which game have you played?

TimeEngineer: UTTT only

MSmits: ohh ok

jacek: not even green circle?

TimeEngineer: And I learned the rules yesterday

MSmits: tons of interesting board games here

TimeEngineer: Among what he says

TimeEngineer: Yes I played GC ofc

TimeEngineer: I used UTTT to learn NN from scratch but I don't really like them :p

MSmits: uttt is not a good starting place for nn afaik

TimeEngineer: What's the best place then ?

jacek: oware

MSmits: oware probably

MSmits: gotta go for a bit, garden work

jacek: its just quite recently NN took over UTTT

TimeEngineer: See you

MSmits: ttyl

jacek: even recurse had problems with them most of the time

TimeEngineer: Why is he so popular ?

MSmits: thats not the right word to use

TimeEngineer: Why even him ?

MSmits: he's just very very good

Illedan: 1 more very needed

MSmits: allright going to pull some weeds now

Illedan: gl

Astrobytes: "weeds" eh

jacek: in holand :thinking:

Astrobytes: :D

Passifi: wooo wooo wooo I did it I finally solved Labyrinth woooo the dopamin rush, the joy .... and its gone :upside_down: next puzzle ;)

jacek: space maze?

eulerscheZahl: bender 4

jacek: galaxy labyrinth?

Passifi: https://www.codingame.com/training/hard/the-labyrinth

eulerscheZahl: we know the puzzle

Passifi: okay I thought there was ambiguity ;)

eulerscheZahl: suggestions for your next one

Passifi: always appreciated yes

eulerscheZahl: there are more mazes to solve

Passifi: oh I see I know see the space maze puzzle in the very hard section I was confused here for a minute and bender 4 is another maze one ?

eulerscheZahl: https://www.codingame.com/multiplayer/optimization/bender---episode-4

eulerscheZahl: there's also sokoban for mazes

Passifi: ah cool thanks

jacek: :tada: https://www.codingame.com/share-replay/650637883

PatrickMcGinnisII: Promotion to Wood 1 League in : 00H 00MN 24SC


jacek: mad pod racing?

PatrickMcGinnisII: CTTT

jacek: ezpz

DomiKo: gz gz

jacek: bz2 even

BlaiseEbuth: Hey hey! use my starter go legend. ezpz

geppoz: Passifi try this https://www.codingame.com/ide/puzzle/the-fall-episode-3

geppoz: well, passing by https://www.codingame.com/ide/puzzle/the-fall-episode-1 and https://www.codingame.com/ide/puzzle/the-fall-episode-2

PatrickMcGinnisII: 63rd, ok

Illedan: Nice jacek. Do you NN consider all inputs together? Or only 1 robot?

jacek: only 1 robot. + number of robots and rounds

Illedan: Nice

jacek: i thought for top5 i would need all of them, but its not needed apparently

j4at: Jacek are you going to try all robots ?

jacek: nah, at least not now

Illedan: Try 2 and 2?

jacek: hm?

Illedan: Give output for all robots seen by 1 robot

Illedan: Then do this for all robots

Illedan: and use the sum of actions as a vote for each of them

Illedan: Then do the best

jacek: oO

Illedan: :P

Illedan: Only 4 moves + defend + 4 attack right?

Illedan: + selfdestruct

jacek: 10 moves

Illedan: 250 outputs then

jacek: quite lately it realized selfdestruct can be useful sometimes

DomiKo: selfplay?

jacek: you could say that, yes

j4at: I'm going to do all robots + rnn :neutral_face:

jacek: thats the spirit

a-Rye: Whoo! I am the tallest midget! Err, dwarf, vertically challenged, whatever the PC version is. I'm top for "Popular" on the Playgrounds lol. Not top coder, not even top explainer, just top popularity of tutorial, most likely accounting for uptime, but I'll take it!

jacek: oO

k4ng0u: a-Rye (not tying to pick up a fight) but whatre are you leading again? I tried to look at a few ranking but never saw you. what is popular tutorial?

jacek: i think he means playgrounds and Popular tab

a-Rye: k4ng0u that's why I joked about being "the tallest midget." I'm ranked fairly average / beginner in most stuff. But my tutorial playground, Beginner Python Concepts, is the most popular. For now, which is why it isn't anything too fancy

a-Rye: https://www.codingame.com/learn?tab=popular

a-Rye: First and Fourth actually, but I'm still working on the Java one. I'm also doing an Intermediate soon, but getting some colleagues to possibly collab on it

jacek: next is Beginner Cobol Concepts

a-Rye: I wouldn't mind learning that. I have a book on Perl I've been meaning to go through as well...

jacek: for you own sanity, dont ;o

a-Rye: Especially just beginner. Anything more, and I'm just posing lol

a-Rye: (╯°□°)╯︵ ┻━┻

a-Rye: A lot of that eh? ^^

Default avatar.png HenroLST: magic!

Westicles: my Stirling numbers clash has a mixed response

Default avatar.png HenroLST: link?

Westicles: https://www.codingame.com/contribute/view/24369cb4a526bd6b707ef8743841a63bd6e67

Westicles: funny that a russian born woman and a soviet born man won wimbledon

Default avatar.png HenroLST: hehe i was hoping krygios would win just to see the upheaval

Default avatar.png HenroLST: also interesting to see if his attitude would change

TomAlard: what happened to recurse's uttt bot

TomAlard: 1->6 lol

DomiKo: MSmits happend

Default avatar.png HenroLST: sheesh

jacek: everything changed when the fire nation attacke

jacek: d

Default avatar.png Evran: hello world

smhb: 👋

Westicles: something changed

jacek: the multiverse is shattering again

Default avatar.png HenroLST: oh no

miszu: question for the code a la mode, what datastructure you used for the string of the dish description?

miszu: Do you guys tokenize it?

MACKEYTH: I would. Minimize string processing that way.

Astrobytes: yeah, tokenized then stored bitwise in an int

MACKEYTH: The discord is still pretty quiet despite the "chat going away" warning.

Default avatar.png Sweshies: are the servers offline?

Default avatar.png hacker535646: hiiiiiiiiiiiiiiiiiiiiiii coders

Default avatar.png BabaYagaCELL: Hello man !

TimeEngineer: Even with a DQN with 90% WR against random, I can't reach silver (UTTT) lmao

Default avatar.png dianissimo_z: Hi all!!!!!!!!!!!!!!!!!!!

Default avatar.png sheeeeeeh: hi sisters !!!!

Default avatar.png Fierceblood: Are all clashes fastest mode?

Default avatar.png Fierceblood: Or am I on a very unlikely streak?

TimeEngineer: There are 3 mode, fastest, shortest and reverse

Default avatar.png Fierceblood: Yeah I know, but I always get fastest for some reason

Default avatar.png Fierceblood: :P

Default avatar.png Fierceblood: Again! XD

Default avatar.png hel-makh: same

Default avatar.png User_xin: Today is another day full of hope

angryPenguin: I got like 6 shortest in a row this afternoon

Default avatar.png TheDoomedClown_155f: Hiiiii

**MACKEYTH hates shortest the mostest

Default avatar.png User_xin: 2022年7月11日10:01:55

Default avatar.png emergenitro: hello

Default avatar.png emergenitro: whats up

MACKEYTH: Playing Code a la Mode

MACKEYTH: Just discovered the source of a whole bunch of forfeits.

MACKEYTH: If your partner times out or otherwise inactivates, their position gets reported as -1,-1.

MACKEYTH: I was getting inactivated due to array out of bounds errors when I tried to plot them on the map.

Default avatar.png Oporekungthai: yo yo yo uits humopdayt

Default avatar.png Oporekungthai: what i need you to do

Default avatar.png Oporekungthai: is tag three big d nig