It’s been a while since I wrote about Iomrascálaí the engine for the game of Go I’m writing in Rust. I will try to do it a bit more often from now on as I’ve finally found the motivation to work on it again.
So today I’d like to announce version 0.3.0! It’s been in the works since September and included two big improvements:
- We’re now using the RAVE heuristic in selecting which tree leaf to investigate next.
- We use a set of 3x3 patterns to guide both the tree exploration and the move selection in the playouts.
These two changes together lead to a strength increase when playing against GnuGo of ~20% on 9x9 and ~25% on 13x13! See the release notes and the change log for detailed listings of what actually changed between 0.2.4 and 0.3.0.
The plan for 0.4
The main goal for 0.4 is to finally get close to equal strength with GnuGo on 19x19. A bit task but where’s the fun in picking easy tasks? ;) To achieve this goal I’m planning to work on the following issues:
- Speed up the playouts! Just 100 playouts per second on 19x19 is really slow and it’s no wonder that the engine has no chance against GnuGo.
- Add criticality to tree selection algorithm. Apparently both Pachi and CrazyStone have had success with adding this as an additional term to the formula.
- Tune the parameters using CLOP. I’ve moved the parameters into a config file so at least technically it’s now easy to run experiments and optimize the parameters.
- Continue searching when the results are unclear. Various engines have had success with searching longer than the allocated time when the best move isn’t clear (i.e. close to the second best move).
- Use larger patterns. Until now the engine only uses 3x3 patterns. It seems worthwhile to investigate if using larger patterns can help.
- Use a DCNN to guide the search. There’s a pre-trained neural network that’s in use by several engines to guide the search and it has improved the results significantly for them. It may be a good idea to investigate this, too.
- The main challenge is computation power! Running 500 games for 9x9 and 13x13 each already takes a few days. And adding 19x19 to the mix will mean that changes will take a long time to benchmark.
- The libraries to efficiently run the DCNN code (like Caffe of Tensorflow) have quite a lot of dependencies and it’s not clear how easily they can be integrated with Rust. It will at least make compiling the bot more difficult for newcomers.
Like I said, quite a challenging plan! But I’m sure it will be a lot of fun. I will leave you with a link to talk by Tobias Pfeiffer about computer Go.