def eval_policy()

in run_offline.py [0:0]

• 19 lines of code
• 4 McCabe index (conditional complexity)

def eval_policy(policy, env_name, seed, eval_episodes=10):
    eval_env = gym.make(env_name)
    eval_env.seed(seed + 100)

    scores = []
    for _ in range(eval_episodes):
        traj_return = 0.
        state, done = eval_env.reset(), False
        while not done:
            action = policy.sample_action(np.array(state))
            state, reward, done, _ = eval_env.step(action)
            traj_return += reward
        scores.append(traj_return)

    avg_reward = np.mean(scores)
    std_reward = np.std(scores)

    normalized_scores = [eval_env.get_normalized_score(s) for s in scores]
    avg_norm_score = eval_env.get_normalized_score(avg_reward)
    std_norm_score = np.std(normalized_scores)

    utils.print_banner(f"Evaluation over {eval_episodes} episodes: {avg_reward:.2f} {avg_norm_score:.2f}")
    return avg_reward, std_reward, avg_norm_score, std_norm_score