Creating a Chess AI with TensorFlow

Learn how to create an AI that plays chess using a convolutional neural network.
Link for the colab page:
colab.research...
Link for the dataset show in the video:
drive.google.c...
More matches:
drive.google.c...
(you can load the pgn at lichess.org/an...)
Note about the dataset:
I tested training the AI on two different datasets, the first Stockfish analyzed the positions up to depth 20. And on the second, depth 0. To my surprise, the AI trained in the second dataset achieved a much better result during the training and during the game, probably because it was easier for AI to learn the instant evaluation of stockfish, if you have another idea why this happened, we can talk about it in the comments below.

[@Philgob]

2:45 note : this is not the "attacked" areas, this is all the squares white can move to, this makes a difference for pawns (can only move up) and kings (can only move to unchecked squares)

[@floowkrunker6610]

Underrated video

[@akhilvashist1339]

@@DigitalSecrets can you reply i have a question to ask

[@DigitalSecrets]

@@akhilvashist1339 Hi, sure! Sorry for the delay.

[@akhilvashist1339]

@@DigitalSecrets Can you create a chess AI using machine learning and feeding it only the games of Mikhail tal and making him use all of Mikhail tal's games as a database and learn from that to mimic his playstyle?

[@Champignon1000]

@@akhilvashist1339 that would be possible yes. if you can find the dataset..

[@xuehuapiaopiao3549]

@@akhilvashist1339 technically yes, but the total of games that mikail tal has played isnt enough for an AI algorithm to learn from.

[@noahbond7179]

Hello! The link to the dataset is no longer working, where did you get the dataset file from originally?

[@scasci3929]

Nice video! About the dataset: It probably works better with depth 0 because a long-term prediction of the game is not required while having a pure intuitive analysis of the current board... so, thanks to that, the network training is more accurate and then, during games, that "long-term prediction" that I mentioned is probably achieved by the minimax algorithm that you implemented to let the ai play.

[@user-dh8oi2mk4f]

this is just not true. Nearly all of the strongest chess engines train their networks on a combination of a low-depth search and the game result(1.0 = win, 0.5 = draw, 0.0 = loss). The only reason that the network trained on depth 20 searches did worse is that depth 20 searches is much too difficult of a training target for a network of that size to learn effectively. If the network were trained on depth 8-9 searches I guarantee it would've performed much better than either of the 2(Also, the testing methodology was very questionable. Looking at the games is not testing, at the very least you should have them play against each other)

[@BPGHchess]

The links don't work anymore.

[@gustavboye6691]

By the way, you can change def minimax_eval(board): board3d = split_dims(board) board3d = numpy.expand_dims(board3d, 0) return model.predict(board3d)[0][0] to def minimax_eval(board): board3d = split_dims(board) board3d = numpy.expand_dims(board3d, 0) return model.predict(board3d)[0][0] model.predict() is super slow for some reason, model() is about 10 times faster.

[@CyanuricMusic]

I think you mean changing model.predict(board3d)[0][0]) to just be model(board3d)[0][0]). The two examples above are the same..

[@Champignon1000]

@@CyanuricMusic Yes exactly, thats what i meant, sorry.

[@nguyenucvinh-BDCCN]

@@Champignon1000 do you have link for source code and dataset ? If you have, can you send it for me ? thanks a lot

[@Champignon1000]

@@nguyenucvinh-BDCCN hey no sorry i dont. But recently went back to my own version of this project. Id strongly recoment getting stockfish to work so you can generate your own data on the fly. you need a endless stream of data to train on

[@nguyenucvinh-BDCCN]

@@Champignon1000 can you send me your version, pls ? KZbin don't allow to write the gmail, so I put my gmail on my description. I'm beginner, so I really want to learn from it. Thanks

[@soccerstuff]

Great Video, the colab link and the google drive doesn't work, is it possible for you to fix it? Thanks!

[@senyoohene3021]

Excellent video. I'm doing a project on this stuff, and this is really helpful.

[@henry-db]

That's WONDERFUL 🤩🤩🤩 !!! 10/10

[@owenjones6784]

I copy-pasted the above code directly, and used the dataset he linked in the comments. I got the same loss as he did after training, yet the model played HORRIBLY when I tested it with the get_ai_move function. Im confused, and help would be appreciated

[@DigitalSecrets]

Sorry for not making it clear, I made some small improvements to the final model that I did not mention in the video (to make the video not too long). I used "principal variation search" to speed up the searching and concatenated coordinates before the convolutions to help the network encode positions. Let me know if I can clarify any more questions.

[@nguyenucvinh-BDCCN]

@@DigitalSecrets Can you send me a new link for source code and dataset, pls ? the old one is not working now

[@ahmadwahab962]

@@DigitalSecrets Can you send me a new link for source code and dataset too ? the old one is not working

[@migueltutu1]

​@@DigitalSecrets can you send me the link too please? Great video by the way

[@RRPLD_Ambassador]

@@DigitalSecrets The link doesn't work anymore, can you send a new link please?

[@lordseidon9]

The reason for the AI better performance at depth 0 over depth 20 would be because at depth 0 stockfish returns an evaluation of what is happening on the board at that moment while at depth 20 it is returning an estimate if what will happen in the future. Training the AI on calm positions at depth 5 should be enough to make a stable AI.

[@nguyenucvinh-BDCCN]

do you have link for source code and dataset ? If you have, can you send it for me ? thanks a lot

[@lordseidon9]

@@nguyenucvinh-BDCCN No, not yet

[@Champignon1000]

Absolutely love your tutorial, having a fun time trying different architectures. I could really need some help on generating some more data on higher search depth, to further train the model. But I'm kinda struggling to get the stockfish to work.

[@nikosalexoudis8874]

I do too get permision denied. I mean I could continue without it but it would be much better to be able to use stockfish. Did you have any luck from then?

[@Champignon1000]

@@nikosalexoudis8874 Im afraid not :(

[@Muhammad-id6rt]

You play chess too ! My respect

[@ponmuthu..4796]

where can i find the dataset

[@wojciechadamiec8167]

Do i see correctly that you started your AI with min-max depth set to 1.? It would mean that you are actually not using min-max at all, just returning min-max eval value (taken directly from your NN model). Can you tell something about performance of your bot (or rather algorithm itself?): How deep can you dig inside moves tree to get correct move and a reasonable time to wait?

[@cokechan1956]

Very concise explanation, thanks

[@FSstan]

Great video! I see that you use a 64 positions output layer. Could you explain this to me? I understand that the boardsize is 8*8, but how do you know which piece to move there?

[@EK-rw9sw]

i think the output layer is just size 1, to output the score given a board state. then you can use this evaluation inside a minimax algorithm or plug it into stockfish.

[@askurmom3513]

Hello, I have been messing around with different tensorflow chess engines and it seems that your engine is the best in its simplicity (it is very simple and very effective - as we can see on your video). Although when i copy-pasted ur code into my colab with the same dataset and stockfish it plays veeeeery poorly. It's like engine is doing random moves without any sense. Did you do any changes in neural network or dataset in your video?

[@aesop1451]

I would like to know if it were possible to create an AI that plays like a Romantic era player by training it on enough games?

[@anto1756]

I'm doing something similar. I think it's the same code but you need to use your own dataset in npz format

[@tornadoreaper]

fantastic

[@wacimo16]

Very useful 👍

[@subbox1]

Hi, I just wanted to clarify that in your code it states "h3->17" but it should be 12

[@anto1756]

I can’t access the google colab file. Apparently it’s not available. Anyone else having the issue ?

[@BPGHchess]

Yes

[@sahildas.]

yes

[@beaverbuoy3011]

Great Video

[@beaverbuoy3011]

Wow

[@Muhammad-id6rt]

I came from "the dark pursuer" game I want to know how to make a match with somebody else ?

[@Lord_Liverpool]

I keep getting an error when I try to train the model.

[@nikosalexoudis8874]

Did you manage to fix it?

[@Lord_Liverpool]

@@nikosalexoudis8874 Sorry no, I didn't have enough time to fix it and was forced to move on.

[@ponmuthu..4796]

hey do you have source code@@Lord_Liverpool

[@dlep9221]

Hi very interesting, are the links off ?

[@trondamlie5853]

Hi. This is new to me... Could someone give me the code to create the dataset for training? (Using random_board and stockfish eval I guess)

[@Philgob]

i also don’t understand this part… how to create the dataset from the random_board() and stockfish() functions

[@jorgeguzmanjmga]

Which could be the difference if we use reinforcement learning? How can we use that method with the same libraries?

[@HarshKaushik-wv2do]

I think the difference would be that this model isnt able to think much in future(higher depth) but deep reinforcement learning one would be able to

[@topasta7411]

why do i get "No such file or directory: '/content/stockfish': '/content/stockfish'" when i try to "print(stockfish(board, 10))"?

[@MrKoszmaro]

if you are working with a copy of the colaboratory then you need to import stockfish and the dataset yourself

[@topasta7411]

@@MrKoszmaro ty

[@WoodyDataAI]

@@MrKoszmaro Author gives dataset link, but no stockfish link, where is the stockfish?

[@CooperNicolaysen]

I keep getting a graph execution error after using model.fit(x_train, y_train)

[@Lord_Liverpool]

I got the exact same error, did you manage to fix it?

[@CooperNicolaysen]

@@Lord_Liverpool Yes. If you’re using newer versions of all these APIs, then the dimensions are in different spaces. So remove the data_format= whatever part of the code where you write 14,8,8,8.

[@Lord_Liverpool]

@@CooperNicolaysen Thanks for the fast reply. OK, great I will investigate your advice, thanks very much!

[@CooperNicolaysen]

@@Lord_Liverpool If you don’t end up getting it, i’ll send you a copy of my working code.

[@Lord_Liverpool]

​@@CooperNicolaysen KZbin won't let me post a comment containing my address. If you click my name and go to the about section of my channel there's a button to reveal my address. Thank you very much for your help, I'm very grateful.

[@davidnguyen6724]

the files are deleted bro plz update it

[@RaushanBharadwaj]

Can you send me the link of dataset

[@midasscheffers7610]

How did you use stockfish in your program, im trying to make a chass bot aswell but that part I can't figure out

[@trondamlie5853]

Just duwnload stockfish from their site and unpack it to your python-folder... and the code works... (using relative path)

[@WoodyDataAI]

@@trondamlie5853 can you give me the site for stockfish? thanks

[@theterror2895]

Good!

[@nikhilmarepally9341]

Could you upload the whole code, I see you implemented 'principla variation search' and encoding could you upload that ?

[@walney2008]

links off, download please, convert png images in FEN ?

[@shivmu1642]

where did you get the dataset from?

[@shaynegoldstein1329]

its in the description

[@iuexwo]

I like your horror gamem❤

[@JohnDuthie]

Can ChatGPT write this for me?

[@forge8807]

You should try learning it yourself, its significantly more rewarding than just using an ai that clearly cant do it

[@theterrorconsumes9021]

this video was interesting! (;

[@protibadimon4169]

I need help can you help me? I need chess developer

[@lordseidon9]

Do you still need now?

[@rotrot.graddiv.-.3766]

Hey, it is extremly interesting ! However I doubt that python is adapted for a performant min/max algorithm. Of course you can easily get to depth 4/5 but I don't think that you can go further without having to wait for a while. I will look if I can extract this neural network to a C or C++ program.

[@TayyabTV]

the gains in C/C++ would be marginal imo, you would have to improve the algorithm as a whole to get better computation.

[@rotrot.graddiv.-.3766]

@@TayyabTV Well I don't think it is ! Python is much much slower than a compiled language. And when I mean slower, it can do from 10 to 100 times slower in some cases. One of other the reasons for that is that on a C/C++ program you can use multiple cores of your CPU while on a Python program you are limited to only one core. Let's pretend that you have 12 cores (the standard for nowdays CPUs), not even considering the execution, you have access to 12 times less power ! And believe me even on a single core this is not even close. C is just way faster. But for the AI thing however you are totaly right since TensorFlow is built on C++ such as most of the Python libraries. In fact Python is initialy meant to make the link between performant libraries by an easy way. Python is slow but the libraries are fast. This is part of the reasons why Numpy is so popular even to just create lists. As a good developper, you have to use the right tools depending on the problem you are facing. You wanna make a quick script that doesn't require much performances ? Python is probably one of the best options ! You wanna build a performant program that does an enormous amount of operations ? You have to use a compiled language, mostly C/C++ or any C-built on language.

[@machamemes3766]

Algum br