Arimaa Forum - Print Page


    
      
        Arimaa Forum
        (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi)
      

        Arimaa >> Bot Development >> Arimaa AlphaZero Challenge
        
(Message started by: omar on Aug 2^nd, 2020, 11:28am)

Title: Arimaa AlphaZero Challenge
Post by omar on Aug 2^nd, 2020, 11:28am

I had posted the following in another thread.

"A year after defeating the top Go player, the Google DeepMind project is continuing to make astonishing progress. Now they don't even need to use a database of expert games to train the AI, they just give it the rules and let it play itself to produce super human AI players. And not just for one game. The same AI can play Chess, Go and Shogi. It masters these games in a matter of hours.

https://goo.gl/c8RWx3

https://arxiv.org/pdf/1712.01815.pdf

If anyone is interested in reproducing these results for Arimaa and making the bot and bot generation code open source (along with a published paper describing the work), I would be willing to award a prize of $10,000. The program has to win a 10 game match against bot_Sharp2015CC to prove superior human level play."

http://arimaa.com/arimaa/forum/cgi/YaBB.cgi?board=other;action=display;num=1512935468

I know RightfulChip is working on this. Are there others developing an AlphaZero type bot that also plays Arimaa?

I'm contemplating how to run a challenge match in 2021. Current thinking is that If there are multiple bots that win a 10 game match against Sharp2015CC then we run a floating triple elimination tournament among these bots to decide the winner of the challenge. Comments and suggestions welcome.

Title: Re: Arimaa AlphaZero Challenge
Post by RightfulChip on Aug 3^rd, 2020, 1:21pm

bot_rusty_zero is a reproduction of AlphaZero and is ready to compete in the 2021 Challenge Match!

Title: Re: Arimaa AlphaZero Challenge
Post by Max2 on Aug 9^th, 2020, 4:40am

It should be defined, how "zero" an "Arimaa AlphaZero Challenge"-bot must be.

In my opinion optimizations should be allowed, e.g.
the use of endgame tablebases (dynamic Win-in-x detection as used by bot_Sharp or fixed computed ones like used by LelaChessZero);
the use of setup/opening tablebases (they also can avoid identical games);
using differnt input planes for every cat (like differnt input planes for white/black bishop in chess);
feeding the neural net additional informations like "is the elephant blockaded"/"is a piece in danger of being caught";
...

If optimizations are ok, a definition for a bot might be then:
"All bot optimizations are allowed.
'zero' here only means:
The bot has to use a neural net, which is trained with zero human/other-bot games, to determine the value of a position. Also every other net has to be trained only on self-play games."

If no or only some optimizations were allowed, there would be a big discussion about every aspect of whether a bot is "zero" enough for the price money, which starts with the time management and ends with the use of history planes (info about the last moves) that even AlphaZero used because it made it much stronger.

If there were more demands,
e.g. also a policy neural net has to be used; monte carlo search has to be used;...
bots like StockfishNNUE in chess (without a policy net and with an alpha-beta serach) would not be allowed to participate.

As a spectator it would also be interesting for the tournament to see bots trained on human/other-bot games (like Dolores) playing unofficially against all "Arimaa AlphaZero Challenge"-bots.

Title: Re: Arimaa AlphaZero Challenge
Post by omar on Aug 9^th, 2020, 12:10pm

The main theme of this challenge is to produce a bot that can learn to play Arimaa from self play and zero external knowledge. Now that Arimaa is not a brand new game, there is a lot of game knowledge floating around which makes this difficult.

But to see what my intent is with this challenge imagine that you build the zero knowledge bot for Arimaa and it defeats Sharp2015 to be a finalist. Now instead of the bots playing Arimaa in the final tournament they will play a variant of Arimaa. The variant may be on a 12x12 board with more pieces and weakest piece being able to push the strongest piece. Your bot will have 1 month to learn this variant. You will only provide the code and setup that I need to run to train your bot and I will run the training on rented hardware. This is not what we are going to do, but it is to illustrate what mean by having zero knowledge about the game it learns to play. In this case, expert games or opening/end game tables would not be available. How you choose to represent the game or what hand coded features you decide to provide may help.

In the actual challenge, I am not going to check how you trained your bot or if is strictly trained with zero external knowledge. So pretty much anyone can make the claim it is a zero knowledge bot and enter the challenge. For me the only concern is if too many bots entered the challenge and time it would take to run the challenge. If we have more than 4 bots winning the match against Sharp2015 I might allow the community to run a pre-tournament to reduce the number of bots to 4.

Also aaaa suggested using a quad elimination tournament since we used that in the previous WCC tournaments. So I'm ok with doing that.

Title: Re: Arimaa AlphaZero Challenge
Post by Max2 on Aug 16^th, 2020, 12:35pm

I am trying to get into practical machine learning. So I will use bot_Maxy, my very weak arimaa playing Java bot, and add machine learning capabilities to it.
I will start with learning from human and bot games. There will be a focus on handicap games. This bot will be called bot_Ananas.
If bot_Ananas is successful and there is enough time till the AlphaZero Challenge, I will also make a complete reinforcement run from scratch for the participation in this challenge calling this bot then bot_Annaa.

Title: Re: Arimaa AlphaZero Challenge
Post by ajzaff on Sep 9^th, 2020, 5:02pm

bot_zoo may be competing if I stick to it.

Title: Re: Arimaa AlphaZero Challenge
Post by Max2 on Nov 1^st, 2021, 8:27am

I have just stopped developing my bots bot_Ananas and bot_Annaa, because there was not much progress during the last months. There might be still too many bugs or other problems.

So I will not participate in the Arimaa AlphaZero Challenge.

Good luck to all other developers and theire bots.

Thank you very much Omar, for this new challenge. It motivated me, to learn more about AI and get into practical machine learning.

Title: Re: Arimaa AlphaZero Challenge
Post by Animiral on Nov 8^th, 2024, 5:24am

Is this prize still on? Because for this amount, it sounds definitely doable.

Since AlphaGo Zero, several improvements to the network architecture and learning process have been invented. Just by applying what has been done in KataGo for Go should probably be enough.