Overview
Javelin is a second UCI chess engine I made. It uses Monte Carlo Tree Search (MCTS) to find best moves. I made it as an experiment, wanting to try something different with MCTS. Javelin works with any chess GUI that supports UCI. Data for training value and policy neural networks was generated through entirely through selfplay from the beginning, when Javelin was using basic PeSTO.
Compiling
To compile Javelin, follow these steps after downloading the source code from the release:
Open folder Javelin-X.X.X
Inside this folder create directory target and inside this directory create another one called builds
Go back to root directory and start a terminal
Start a terminal in root folder and enter the make command
Author: Tomasz Jaworski
v.2.0.0:
What's Changed
Changed value net size from 768->16->1 to 768->128->1
Added horizontal mirroring to value net based on side to move king file
Changes policy net architecture from 768->384 to 128 subnet pairs (768->16)
Added selection of a policy subnet based on the result of SEE
Increased inference speed
Increased pulling value from engine options
Preallocating search tree vectors and new nodes vectors
First play urgency
Updated default engine options
Added scaling C value with search duration
Fixed termination bug in tree replacement
Full Changelog: 1.0.0...2.0.0
Estimated Elo: 2539
Chess engine: Javelin 2.0.0 download
Comments
Post a Comment