def __init__(self,
config,
policy,
global_episode,
n_agent=0,
agent_type='exploration',
log_dir=''):
print(f"Initializing agent {n_agent}...")
self.config = config
self.n_agent = n_agent
self.agent_type = agent_type
self.max_steps = config['max_ep_length']
self.num_episode_save = config['num_episode_save']
self.global_episode = global_episode
self.local_episode = 0
self.log_dir = log_dir
# Create environment
self.env_wrapper = create_env_wrapper(config)
self.ou_noise = OUNoise(dim=config["action_dim"],
low=config["action_low"],
high=config["action_high"])
self.ou_noise.reset()
self.actor = policy
print("Agent ", n_agent, self.actor.device)
# Logger
log_path = f"{log_dir}/agent-{n_agent}"
self.logger = Logger(log_path)
def eval_policy(self, eval_episodes=10):
env_wrapper = create_env_wrapper(self.config)
avg_reward = 0
for _ in range(eval_episodes):
state = env_wrapper.reset()
done = False
while not done:
action = self.target_policy_net.get_action(
state).detach().cpu().numpy().flatten()
next_state, reward, done = env_wrapper.step(action)
avg_reward += reward
state = next_state
if done:
break
avg_reward /= eval_episodes
print("---------------------------------------")
print(f"Evaluation over {eval_episodes} episodes: {avg_reward:.3f}")
print("---------------------------------------")
return avg_reward