Reinforcement Learning Chess

Parameters

Parameter	Type	Description	Example Values
n_episodes	int	Number of episodes to train	100
state_dim	int	Dimension of states	64
action_space	int	Dimension of action space	4096
k_steps	int	Number of unroll steps for training	5
n_simulations	int	Number of MCTS simulations per move	800
max_game_length	int	Maximum length of a game	1000
c_puct	float	PUCT exploration constant for MCTS	1.5
batch_size	int	Training batch size	8
learning_rate	float	Learning rate for optimizer	1e-3
seed	int	Random seed for reproducibility	2025
noise	bool	Add Dirichlet noise to root node priors	True/False
noise_alpha	float	Dirichlet noise alpha parameter (concentration)	0.3
noise_epsilon	float	Dirichlet noise epsilon parameter (mixing ratio)	0.25
save_every	int	Save model checkpoint every n steps	10000
verbose	bool	Enable verbose output during training	True/False

Train Model

python train.py --noise --n_episodes 1000000

Test Model

python test.py