def _run_agent_one_ep(env: BaseEnv,
agent: DQNAgent,
config: Config,
eps: float,
behavior_name: str,
train: Optional[bool] = True):
# Get a starting state
env.reset()
decision_steps, terminal_steps = env.get_steps(behavior_name)
state = decision_steps.obs[0]
agent_id = decision_steps.agent_id[0]
done = False
did_win = False
episode_reward = 0.0
import time
while not done:
reward = 0.0
# Get and perform an action
action = agent.act(decision_steps.obs[0], eps)
env.set_actions(behavior_name,
np.expand_dims(action, 0).reshape(-1, 1))
env.step()
decision_steps, terminal_steps = env.get_steps(behavior_name)
# Determine S', R, Done
next_state = None
if agent_id in decision_steps:
reward += decision_steps.reward[0]
next_state = decision_steps.obs[0]
if agent_id in terminal_steps:
terminal_reward = terminal_steps.reward[0]
# Add win/loss
did_win = True if math.isclose(terminal_reward, 1.0) else False
reward += terminal_reward
next_state = terminal_steps.obs[0]
done = True
assert next_state is not None, f"next_state cannot be None. Agent {agent_id} did not appear in decision or terminal steps"
if train:
# Learn from (S, A, R, S')
experience = Experience(state, action, reward, next_state, done)
agent.step(experience)
# Set new state
state = next_state
episode_reward += reward
return (episode_reward, did_win)
