WeiqiDMC stands for Weiqi Distributed Monte Carlo, just a fancy temporary name because naming it GoSomething would have made matters confusing, wouldn't it?
As it stands, the project is just a pet-project made to test out the hypothesis that performance losses in Elixir (compared to C/C++, especially when it comes to memory management) can be offseted by the ease of writing distributed applications.
Supported GTP (Go Text Protocol) commands:
[:version, :name, :protocol_version, :list_commands, :quit, :known_command, :boardsize, :clear_board, :komi, :fixed_handicap, :play, :genmove, :showboard, :set_game]
Current engine is a Random Player. It has no other purpose than test out the rest of the framework.
Installation
For profiling, you'll need to run mix deps.get, there are no other dependencies ATM.
With GTP front-end
Many GTP front-end seem particularity finicky with the path to the binary, so I ended-up just writing a simple bash script (chmod +x) and use the path to that script:
#!/bin/bash
cd FULL_PATH/TO/PROJECT && MIX_ENV=prod FULL_PATH/TO/MIX/BINARY run -e "WeiqiDMC.start_gtp_server"`
I used Quarry to test it out. Don't forget to run MIX_ENV=prod mix compile first :)
The tests
Same as normal, just mix test
- Implement MC + RAVE
- Copy the imperative algorithm from Sylvain Gelly's 2011 paper (MCTS and RAVE Estimation in Computer Go) with Qeven heuristic
- Write tests to secure the implementation (WIP)
- Tune it to get a basic level of strength on 9x9, say 5-10k, with fixed time settings
- Write one of the better heuristic proposed by Sylvain
- Tune it to get it to an intermediate level of strength on 9x9, 2k-1d
- Make it distributed by spawing Erlang processes to perform multiple simulation in parallel, then regroup to calculate the new Q, Qtilde, N and Ntilde
- Add time management
- Add pondering
- Devise a protocol for testing vs other engines
- Tune the engine
- Glory
- Resign from the project and start a fork to make things right
Please do!