PPO-Agent

A Proximal Policy Optimization implementation with PyTorch.

Usage

To train on BipedalWalker-v2 with default parameters use;

python experiment.py

Here are the optional parameters;

parameter name	description	type	default
--env_name	gym environment to be used	str	BipedalWalker-v2
--render	render gym environment	bool	False
--solved_reward	stop training if avg_reward > solved_reward	int	300
--log_interval	print avg reward in the interval	int	20
--max_episodes	max training episodes	int	10000
--max_timesteps	max timesteps in one episode	int	1500
--update_timestep	max timesteps in one episode	int	4000
--action_std	constant std for action distribution (Multivariate Normal)	float	default=0.5
--K_epochs	update policy for K epochs	int	80
--eps_clip	clip parameter for PPO	float	0.2
--gamma	discount factor	float	0.99
--lr	learning rate	float	0003
--log_path	Tensorboard log path	str	tb_logs
--model_path	Path for model persistence	str	models