85 points by open_chess 4 months ago flag hide 14 comments
john_doe 4 months ago next
Interesting project! I've been following the development of AI in games and it's amazing how far we've come. I'd be curious to see how this engine performs against other open-source chess engines.
ai_enthusiast 4 months ago next
@john_doe I agree! It's definitely exciting to see the advancements in this field. I'll be looking forward to benchmarking its performance in the near future.
random_user 4 months ago prev next
What libraries and frameworks were used for this project?
op 4 months ago next
@random_user The majority of the machine learning components were implemented using TensorFlow and Keras. We also utilized the popular chess library 'chess.py' for parsing and performing chess-related functions.
master_player 4 months ago prev next
I noticed that the project utilizes a convolutional neural network. As far as I know, CNNs aren't typical for game tree search algorithms. How well does it scale when compared to more conventional approaches using NNs?
op 4 months ago next
@master_player Good point. A more conventional NN would be a recurrent NN. In this case, the CNN allowed us to have a slightly different perspective and it requires less training data. The performance was promising, but we will continue testing and improving the solution.
stats_junkie 4 months ago prev next
What were the win rates with different ELO rated bots throughout the training process?
op 4 months ago next
@stats_junkie This is a great question! We did track these stats throughout the training and validated the performance across a wide range of Elo ratings. At the end, our engine maintained a win rate of approximately 85% against 2300-rated bots, and a draw rate of about 13% with 2600-2700-rated bots.
chess_developer 4 months ago prev next
Very cool! I've been working on a chess engine as well, but had not thought of leveraging a CNN. This opens new doors and opportunities. Any chance you can make this open source as well?
op 4 months ago next
@chess_developer Thank you! We're glad to have inspired you. The source code has already been open sourced. You can find the URL in the project's GitHub repo linked at the top of the submission.
anonymous 4 months ago prev next
Any plans to expand this to other games beyond chess?
op 4 months ago next
@anonymous Currently, our focus is on mastering the game of chess and taking the engine to the next level. However, there might be opportunities for expansion once this project reaches its full potential.
critical_thinker 4 months ago prev next
Have you considered more advanced training techniques such as AlphaZero style self-play or leveraging a GAN or RL for training?
op 4 months ago next
@critical_thinker Yes, those ideas were on the table and we think they are fascinating directions to explore. During initial testing, we discovered that these techniques required more computational resources than our current setup allowed. Going forward, we intend to implement, test, and benchmark these ideas on larger infrastructure.