def score(df: pd.DataFrame) -> float:
"""Output a score for MDP Playground."""
df = mdpp_preprocess(df_in=df)
regret_score = plotting.ave_regret_score(
df, baseline_regret=BASE_REGRET, episode=NUM_EPISODES)
norm_score = 1.0 * regret_score # 2.5 was heuristically chosen value to get Sonnet DQN to score approx. 0.75, so that better algorithms like Rainbow can get score close to 1. With a bigger NN this would mean an unclipped score of 1.1 for Sonnet DQN, which is fair I think. However, a2c_rnn even reached 2.0 on this scale. DQN may be not performing as well because its epsilon is not annealed to 0.
# print("unclipped score:", norm_score)
norm_score = np.clip(norm_score, 0, 1)
return norm_score
def score(df: pd.DataFrame) -> float:
"""Output a single score for cartpole = 50% regret, 50% has a good run."""
cp_df = cartpole_preprocess(df_in=df)
regret_score = plotting.ave_regret_score(
cp_df, baseline_regret=BASE_REGRET, episode=NUM_EPISODES)
# Give 50% of score if your "best" episode > GOOD_EPISODE threshold.
solve_score = np.mean(
cp_df.groupby('bsuite_id')['best_episode'].max() > GOOD_EPISODE)
return 0.5 * (regret_score + solve_score)
def score(df: pd.DataFrame) -> float:
"""Output a single score = 50% regret, 50% "final accuracy"."""
regret_score = plotting.ave_regret_score(
df, baseline_regret=BASE_REGRET, episode=sweep.NUM_EPISODES)
final_df = df.copy()
final_df['ave_return'] = (
1.0 - (final_df.total_regret.diff() / final_df.episode.diff()))
final_df = final_df[final_df.episode > 0.9 * NUM_EPISODES]
# Convert (+1, -1) average return --> (+1, 0) accuracy score
acc_score = np.mean(final_df.ave_return + 1) * 0.5
return 0.5 * (regret_score + acc_score)
def score(df: pd.DataFrame) -> float:
"""Output a single score for swingup = 50% regret, 50% does a swingup."""
df = cp_swingup_preprocess(df_in=df)
scores = []
for _, sub_df in df.groupby('height_threshold'):
regret_score = plotting.ave_regret_score(
sub_df,
baseline_regret=BASE_REGRET,
episode=NUM_EPISODES,
regret_column='perfection_regret')
swingup_score = np.mean(
sub_df.groupby('bsuite_id')['best_episode'].max() > GOOD_EPISODE)
scores.append(0.5 * (regret_score + swingup_score))
return np.mean(scores)
def score(df: pd.DataFrame) -> float:
"""Output a single score for catch."""
return plotting.ave_regret_score(df,
baseline_regret=BASE_REGRET,
episode=sweep.NUM_EPISODES)
def score(df: pd.DataFrame) -> float:
"""Output a single score for mountain car."""
cp_df = mountain_car_preprocess(df_in=df)
return plotting.ave_regret_score(cp_df,
baseline_regret=BASE_REGRET,
episode=sweep.NUM_EPISODES)