Chat:World/2021-07-11

From CG community
Revision as of 05:29, 12 July 2021 by Chat Log (talk | contribs) (Created page with "<img src=/a/17784428403916> MSmits: Marchete are you there? <img src=/a/17784428403916> MSmits: I found another example where you can see mikla does not even finish endgames...")
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

MSmits: Marchete are you there?

MSmits: I found another example where you can see mikla does not even finish endgames properly and here he loses because of it:

MSmits: https://www.codingame.com/replay/569610594

MSmits: (frame 97)

MSmits: he should have just taken all 4 boxes and he would have won 24-25, instead he gives them all to me as he is just using a heuristic to play endgames

MSmits: it almost never matters because control almost always gives you the win, except in close games like this

MSmits: my bot does not adjust it's expected score to 25-24 because it only does so after a search and my bot didn't need one

opitt: Do you know if/where coc puzzles are stored? I want to redo my code but forgot the test cases …

MSmits: you can view them in contributions if you have done enough clashes

MSmits: not sure how many you need

punter147: MSmits what do you use for solving dots and boxes?

punter147: mcts?

MSmits: negamax

MSmits: mcts is good for solving in general, but I don't think it would work very well here

MSmits: you need transpositions

MSmits: ab pruning + TT still makes my head hurt

MSmits: I generally just mess around with it until a 100 million sample produces no errors

punter147: oh i see. transpositions is like a dynamic programming approach, in which you dont calculate the answer to already seen state again and again...

MSmits: yeah, which is easy without ab pruning

MSmits: but ab pruning makes it so that you dont finish the exact calculation but still want to store the result

MSmits: so you use upper and lower limits in the TT

MSmits: which is where my head starts to hurt

punter147: lol if your head would hurt, a lot of things will hurt on my side :joy: . Just kidding, thanks for the tip MSmits i will start studying ab pruning and transposition tables.

MSmits: D&B makes it even weirder because the same board can occur from player 1 and from player 2's perspective

MSmits: https://homepages.cwi.nl/~paulk/theses/Carolus.pdf

MSmits: best source for ab + TT

MSmits: it has good pseudo code

MSmits: i adapted it for D&B though, because of the symmetry I just mentioned

punter147: aah yes D&B does not distinguish among the lines right, unlike chess positions which are distinguished by the color and the rank of the piece...

MSmits: exactly, and not just that, but because you can have multiple lines in 1 turn because of what you choose to do, the same board can occur at turn 56 or at turn 58 and such

MSmits: mmh not sure i said that correctly

MSmits: I should say, it could be player 1's turn or player 2's turn at the same frame and board, depending on what happened before

MSmits: because you get another play if you're the one finishing a box

MSmits: and sometimes you can finish 2 boxes at the same time but get only 1 extra turn, which switches the parity

punter147: yes you are right MSmits, i was just looking at D&B plays, once you finish a box, you get another chance, again if you make another box, another chance and the party continues...

MSmits: exactly, it's quite messy to code a bot for but mathematically really interesting

MSmits: if you want to practice negamax and such, dont do it here really

MSmits: i mean not in D&B

Marchete: are you using minimax?

MSmits: negamax yes, same thing i guess

MSmits: with ab/tt

punter147: MSmits oh, are there any games simple enough to get our feet wet in negamax?

MSmits: pre-end game a move is a line, in the end game, a move is offering a full string/chain or when offered, it's the choice between doubledealing or not

MSmits: punter147 yes many

MSmits: find a simple board game you know you can simulate properly

MSmits: not sure which is easiest, people tend to go for ultimate tic tac toe even though mcts works better there

Marchete: I mean not MCTS

MSmits: nope not mcts

MSmits: TT is extremely strong here because of the colorless board

MSmits: using TT with mcts is not very effective most of the time

Marchete: do you think he just randomly put lines until a certain moment?

MSmits: I think he has a chaincounting negamax running until he solves the chaincount result

MSmits: and if he doesnt solve, he randoms

Marchete: so no avoiding complex chains?

Marchete: or loops or whatever

MSmits: I don't think so. Avoiding them doesn't really help either player specifically. It just makes the game easier to solve

MSmits: but that goes for both sides

MSmits: I have been considering doing a complexity mcts

MSmits: where i try to get the most complex connection of short chains and such in the end game

MSmits: so that mikla cant deal with it :)

MSmits: it would lead to funny endgames i think, but its a lot of work

MSmits: I am trying to get those sacrifices into my solver now

MSmits: hard to detect them without building a graph

MSmits: i think it can be done with bitboards but not easily

Marchete: early game give bonus for number of chains

Marchete: or whatever target the enemy can't deal

Marchete: I'd say daisy chains are hard

Marchete: or whatever it was called

Marchete: 3-way loop + ground chain

MSmits: well that makes sense, but it's not the current number of chains that matter, it's the amount at the end of the game

MSmits: so if you have half finished structures and empty spaces, you're not gonna know how many chains are gonna be there

Marchete: if you play with that target

Marchete: more than with pure random

MSmits: you can do mcts and randomly rollout until the end game happens, then analyze

MSmits: but thats what i was planning to do already :)

MSmits: except with negamax

MSmits: but i might do it with mcts later. I just don't think it'll be better

MSmits: the problem is that there is no neat progression from a "bad state" to a "good state" by several good moves

MSmits: they are separated by 1 wrong line

MSmits: so my sense is that you need a full solve to make any prediction

MSmits: in games like chess and such you can make your boardstate progressively better or worse with good or bad moves. That's just not the case here

MSmits: it's a coin you can flip between loss and win, except you have no idea which way it is facing before you flip. So it becomes in effect a totally random coin flip

MSmits: I suppose mcts might still give *some* improvement if you pick moves that are more likely to lead to a beneficial flip

Marchete: I was thinking like I said

Marchete: trying to have better patters (like chain types and similar)

Marchete: and try to reach them

Marchete: but as you say

Marchete: going from win to loss is really near

Marchete: but I still think

Marchete: if enemies can't handle daisy chains easily

Marchete: it will be harder for them to solve

MSmits: i dont think you should call them daisy chains

Marchete: I can't remember its name

MSmits: no, thats fine, but you'll get jacek all excited with that term :P

jacek: shackles

Marchete: :D

MSmits: dont you just mean loops?

Marchete: I mean a loop

MSmits: ah ok

Marchete: with a ground chain attached

Marchete: like a flower

MSmits: oh that's called a dipper in literature

Marchete: ahh

MSmits: you always capture the attached chain first

MSmits: then it becomes a loop

Marchete: yes

MSmits: doesnt have to be grounded either

MSmits: the string can be connected to something else

MSmits: it will still be better to do the string before the loop

MSmits: there's a lot of little rules like that that makes the solving faster if you put them in

Marchete: but you need to detect it

MSmits: yes

MSmits: if (joins[0] == joins[1]) { if (joins[0]->ActiveLinks() == 3) return false;

MSmits: joins[0] == joins[1] finds a string that is connected to the same join ( a loop)

MSmits: if the join happens to have exactly 3 strings connected, then it's a dipper

MSmits: so i disallow the doubly connected string to be played in that case

Default avatar.png swapbee: hello

Marchete: but all the 3-way chains are hard to evaluate, and maybe promoting it in the eval is positive

Marchete: as the enemy will have a harder time to eval

MSmits: they are hard to solve yes

MSmits: but anything like that will hurt both sides

MSmits: so this only works if you know you are especially good at dealing with it

Marchete: one will be hurt harder

MSmits: and everyone else isnt

Marchete: exactly

MSmits: it's a good point

Marchete: anyways everyone else seems to mean 1 guy

Marchete: maybe 2 when I tackle the game

MSmits: in this case yeah, he's the only one doing the chain counting efficiently

MSmits: though some other guys seem to make sacrifices really short before the end game (2-3 plays) where he does them 15 plays before the end game

MSmits: my leaderboard version also makes sacrifices really short before the end game

MSmits: thats why i am 2nd really, not because my solver is good

Marchete: remi is near you in score

Marchete: so it must be good too

MSmits: yeah, but he makes sacrifices too easily, even when he's not supposed to. This costs him boxes

MSmits: I think he might be using a chain-counting mcts like we discussed

MSmits: but if you havent fully solved, you should not sacrifice

jacek: join #ru and ask him directly

MSmits: remi is french :P

jacek: oh i thought you meant milka

MSmits: ahh no, i dont need to ask really. I kinda know what he does. He wasnt really secretive about it either

MSmits: https://www.codingame.com/replay/569616446 here you see remi do sacrifice at frame 25 even

MSmits: there is no way he knows it will help

MSmits: and he loses the game so...

Marchete: too early it seems

jacek: premature sacrifice eh

MSmits: exactly :P

Marchete: I'll ad a damn NN and let it do the guessing :D

Marchete: add*

MSmits: yeah for early game this might work

MSmits: or at least, give some small edge

Marchete: and endgame, damn too :D

MSmits: nn will also have trouble converging on a good set of weights, but it only has to be better than nothing

Marchete: just throw more games

MSmits: yeah, but if the data seems too random to it, it wont help

MSmits: your training loss will get smaller but the test/validation loss wont change

Marchete: win/loss will be brought back from endgame

MSmits: but as i said, it only has to generalize just a teeny bit

Marchete: like jace_k does

Marchete: but jk

Marchete: the game is huge in search space

MSmits: yeah

MSmits: I was thinking of trying my solver on smaller versions of D&B to see how it stacks up against the solvers in those papers

MSmits: up to 5x5 has been solved

MSmits: the increase in time required goes up extremely fast with 1 size added

HassanAli4651: can any one help me with this?

MSmits: so 7x7 is miles away, I think even outside of googles range

MSmits: this... what?

HassanAli4651: mars lander ep1

MSmits: simple if/else solution will work

HassanAli4651: i tried

MSmits: if you're going too fast, activate the thrust

MSmits: otherwise just fall

MSmits: dont do any other movement

Marchete: I want someday to try AlphaZero but with graph/features and not the board representation directly, and possible moves as something of high level (like cut a dipper)

jacek: graph game eh

MSmits: seems hard Marchete

MSmits: but many things worth doing are

Marchete: game -> abstraction->AZ->high level move

MSmits: you need a lot of output though, the game can have many suchs constructs and in many combinations

MSmits: not sure how to code the policy here

Marchete: I think less than x,y,side output

MSmits: I think there are thousands of possible objects

Marchete: identify possible chains, and possible cuts inside them

Marchete: and 1 random

Marchete: thousands types->

Marchete: on the same type, it will be a best to cut

MSmits: just looking at a dipper for example. It matters how long the loop type is and also how long the attached chain is

Marchete: by size and similar

MSmits: so you have like 10 x 10 options right there

MSmits: for differen dippers

Marchete: that's true

MSmits: and that does not include ungrounded dippers

MSmits: those will have a massive amount of variation

MSmits: I think x, y side might be easier to do :)

MSmits: btw, I made a lookup table for simple endgames

MSmits: so just isolated chains and loops

MSmits: I only needed 6k values

MSmits: but using the lookup table wasnt faster than just running the function

MSmits: (the one i shared)

MSmits: that error by mikla's bot i shared earlier would have been 1 of the lookup values

Default avatar.png SlickSteel: Hey yall!

jacek: ohai

sakuraiuto001: hey

CodeAce: ?

Marchete: ¿

doomento: hi

Plantchant: hello

Default avatar.png client2233: hello

mokupnik: Is there some issue on the site? Because i can't make a contribution, building just goes on forever. Even when I upload he project that was already accepted.

jacek: try again

Default avatar.png binjn: hello everyone

walker2k: hello binjn

walker2k: :D

Default avatar.png binjn: oh hello

Default avatar.png FriendlyKh: y

Default avatar.png TheIceJetski_9b4a: hello

Konrad_O: Hello!

AntiSquid: oh look CSB with lazors

jacek: ?

Default avatar.png QuickMathzs: :poop:

BugKiller_328: hi

BugKiller_328: anyone here now ?

BlaiseEbuth: No.

BugKiller_328: :smile:

BugKiller_328: in 'PRACTICE' puzzles, (easy, medium, hard), is there any way to search puzzles which has specific learning opportunity ?

BugKiller_328: for example, search problems which have 'backtracking' as its 'LEARNING OPPORTUNITY'

BlaiseEbuth: Not from this location. But you can click the 'tag' corresponding to what you want to practice to get a list of all the puzzles that have it.

BugKiller_328: cool, clicking tag will list me the puzzles associated with it ?

BugKiller_328: Thanks..

BlaiseEbuth: yup

BugKiller_328: any location to display list of tags ?

BlaiseEbuth: And if you want a full list of tags with the puzzles associated check https://chadok.info/codingame/tags_list.html

BlaiseEbuth: ^^

BugKiller_328: yes, this is what I want exactly.

BugKiller_328: I'm spending much time here these days to learn & have fun..

BugKiller_328: @BlaiseEbuth thanks.

BlaiseEbuth: :thumbsup:

BlaiseEbuth: THe @ isn't required btw BugKiller_328

BugKiller_328: haha.. this is also I didn't know. just write name and it will work ? BlaiseEbuth test.

BlaiseEbuth: Yup that works

BugKiller_328: my test message is displayed in red on your side ?

BlaiseEbuth: Yes

BugKiller_328: in my side, no any diff

BugKiller_328: ok. thanks. (maybe platform highlights the text which has my nickname )

jacek: also when you write your password it will be masked automatically. like ********

BlaiseEbuth: Same for your bank card code

Counterbalance: wtf https://www.codingame.com/learn/%F0%9F%92%A1%F0%9F%92%A1%F0%9F%92%A1%F0%9F%92%A1%F0%9F%92%A1

Counterbalance: emojis in urls.. ieuw...

BlaiseEbuth: https://www.emojicode.org/

BugKiller_328: :rolling_eyes:

jacek: https://img-9gag-fun.9cache.com/photo/aoMq21A_460svav1.mp4

BlaiseEbuth: :D

jacek: i wonder if that game would fit in CG

BlaiseEbuth: Only one command "play"... The branching will not be too important.

jacek: you can choose pawn

BlaiseEbuth: Oh my! Oo

BlaiseEbuth: Simulation will be complicated though with the random...

BlaiseEbuth: How to transform a quick fix into a big one: Correct an old version of the code, then copy-paste it over the actual code, and finaly, commit. Good. You can now fix your bullshits.

Westicles: Do we know the date when chat is going to die?

BlaiseEbuth: There's nothing planned about that.

Westicles: Ah, good to hear

BlaiseEbuth: CG don't want to spend time updating/recoding the chat, so they spoke about its suppression. But some of the CG devs are against it, so for now it will just stay as it.

BlaiseEbuth: Why this question btw?

Astrobytes: 2061 it will die

Westicles: Us non-mods just have to try and read the tea leaves

Astrobytes: We get no special info for the most part unless there's a specific issue

Astrobytes: To their credit, CG are pretty transparent in that regard

pci: Sokoban -- I finally figured out what I was doing wrong and why I was struggling to solve some of the test cases within 10 seconds...

pci: the way I was encoding the state for my visited test was preserving the order of the boxes... by removing that suddenly my 15 second times turned to 1.5 seconds

pci: Unfortunately I tried a lot of different algorithms before discovering this... then went back to my original and that was all that was needed. :P

Lordam: Hi all, its very embarassing but even after reading the forum about the puzzle "Encryption/Decryption of Enigma Machine" I cannot figure out why the test cases 4 and 5 always fail (ingame all test cases are green). Anyone any hint? :S

Lordam: got some help from Westicles already, thanks!!

Wontonimo: what? world chat could go away?

Wontonimo: just getting caught up on the chat from hours ago

Default avatar.png ElderMan: there isn't any bare bones tutorials on here, is there?

Default avatar.png ElderMan: I'm extremely new.

GarethEddies: I'm looking for someone who can provide some feedback on a code challenge submission before I send it for draft submission

Default avatar.png BobLob: I have some time if you haven't gotten a response yet.

Default avatar.png Eeveee: World Chat could be great way to learn new things from many people, it just doesn't make sense that they are considering to kill it just because it's difficult to maintain

Default avatar.png BobLob: I'm just hearing about this tonight, but it would be rather ironic if a company that was trying to sell its expertise in the coding field found a chat service difficult to maintain.

Default avatar.png BobLob: Or, maybe maintenance isn't the issue? Maybe it's lack of use. There have been 7 messages in the last 6 hours.

juice0: is the discord channel more popular?

Wontonimo: nope, the discord channel is pretty dead

Wontonimo: it gets busy a couple times a week, and very busy during a multi challenge

Wontonimo: it = here on world chat

bitwiseio: im handycaped, im drunken :D

CodeAce: Hello World

CodeAce: In Real World

CodeAce: :joy: