def solve():
env = EnvironmentWrapper(gym.make('LunarLander-v2'))
n_episodes = []
for i in range(1):
returns = dqn_with_prioritized_experience(env, n_episodes=None)
n_episodes.append(len(returns))
if i == 0:
baseline_returns = data_exploration(env,
n_episodes=len(n_episodes))
data.report([(returns, 'b', 'Solution'),
(baseline_returns, 'r', 'Baseline')],
title='Solution',
file='lunarlander_solve_dqn.png')
n_episodes = np.array(n_episodes)
print('LunarLander solved!')
print(' Median: {} episodes'.format(np.median(n_episodes)))
print(' Mean: {} episodes'.format(np.mean(n_episodes)))
print(' Std: {} episodes'.format(np.std(n_episodes)))
print(' Min: {} episodes'.format(np.min(n_episodes)))
print(' Max: {} episodes'.format(np.max(n_episodes)))
print(' % diverged: {}'.format(
len(n_episodes[n_episodes >= 2000]) / float(len(n_episodes))))
def solve():
env = EnvironmentWrapper(gym.make('CartPole-v1'))
n_episodes = []
for i in range(10):
returns = dqn_with_fixed_targets(env, n_episodes=None)
n_episodes.append(len(returns))
if i == 0:
baseline_returns = data_exploration(env,
n_episodes=len(n_episodes))
data.report([(returns, 'b', 'Solution'),
(baseline_returns, 'r', 'Baseline')],
title='Solution',
file='cartpole_solve_fixedq_dqn.png')
n_episodes = np.array(n_episodes)
print('CartPole solved!')
print(' Median: {} episodes'.format(np.median(n_episodes)))
print(' Mean: {} episodes'.format(np.mean(n_episodes)))
print(' Std: {} episodes'.format(np.std(n_episodes)))
print(' Min: {} episodes'.format(np.min(n_episodes)))
print(' Max: {} episodes'.format(np.max(n_episodes)))
print(' % diverged: {}'.format(
len(n_episodes[n_episodes >= 2000]) / float(len(n_episodes))))
''' Sample output:
def run_single_trials():
env = EnvironmentWrapper(gym.make('CartPole-v1'))
n_episodes = 500
baseline_returns = data_exploration(env, n_episodes)
data.report([(baseline_returns, 'b', 'Baseline')],
title='Random Walk',
file='cartpole_single_random_walk.png')
basic_dqn_returns = basic_dqn(env, n_episodes)
data.report([(basic_dqn_returns, 'b', 'Basic DQN'),
(baseline_returns, 'r', 'Baseline')],
title='Vanilla DQN',
file='cartpole_single_basic_dqn.png')
dqn_w_exp_returns = dqn_with_experience(env, n_episodes)
data.report([(dqn_w_exp_returns, 'b', 'DQN w/ ER'),
(baseline_returns, 'r', 'Baseline')],
title='Experience Replay',
file='cartpole_single_er_dqn.png')
dqn_w_fixed_targets_returns = dqn_with_fixed_targets(env, n_episodes)
data.report([(dqn_w_fixed_targets_returns, 'b', 'DQN w/ Fixed-Q'),
(baseline_returns, 'r', 'Baseline')],
title='Fixed-Q Targets',
file='cartpole_single_fixedq_dqn.png')
# Plot all the variations
data.report([(basic_dqn_returns, 'b', 'Basic DQN'),
(dqn_w_exp_returns, 'g', 'DQN w/ ER'),
(dqn_w_fixed_targets_returns, 'm', 'DQN w/ Fixed-Q'),
(baseline_returns, 'r', 'Baseline')],
title='All DQN Variants',
file='cartpole_single_all_dqn.png')
def run_multiple_trials():
env = EnvironmentWrapper(gym.make('LunarLander-v2'))
n_episodes = 500
n_trials = 10
baseline_returns = []
for i in range(n_trials):
baseline_returns.append(data_exploration(env, n_episodes))
data.report([(baseline_returns, 'b', 'Baseline')],
title='Random Walk',
file='lunarlander_multi_random_walk.png')
basic_dqn_returns = []
for i in range(n_trials):
basic_dqn_returns.append(basic_dqn(env, n_episodes))
data.report([(basic_dqn_returns, 'b', 'Basic DQN'),
(baseline_returns, 'r', 'Baseline')],
title='Vanilla DQN',
file='lunarlander_multi_basic_dqn.png')
dqn_w_exp_returns = []
for i in range(n_trials):
dqn_w_exp_returns.append(dqn_with_experience(env, n_episodes))
data.report([(dqn_w_exp_returns, 'b', 'DQN w/ ER'),
(baseline_returns, 'r', 'Baseline')],
title='Experience Replay',
file='lunarlander_multi_er_dqn.png')
dqn_w_fixed_targets_returns = []
for i in range(n_trials):
dqn_w_fixed_targets_returns.append(
dqn_with_fixed_targets(env, n_episodes))
data.report([(dqn_w_fixed_targets_returns, 'b', 'DQN w/ Fixed-Q'),
(baseline_returns, 'r', 'Baseline')],
title='Fixed-Q Targets',
file='lunarlander_multi_fixedq_dqn.png')
dqn_w_per_returns = []
for i in range(n_trials):
dqn_w_per_returns.append(
dqn_with_prioritized_experience(env, n_episodes))
data.report([(dqn_w_per_returns, 'b', 'DQN w/ PER'),
(baseline_returns, 'r', 'Baseline')],
title='Prioritized ER',
file='lunarlander_multi_per_dqn.png')
# Plot all the variations
data.report([(basic_dqn_returns, 'b', 'Basic DQN'),
(dqn_w_exp_returns, 'g', 'DQN w/ ER'),
(dqn_w_fixed_targets_returns, 'm', 'DQN w/ Fixed-Q'),
(dqn_w_per_returns, 'c', 'DQN w/ PER'),
(baseline_returns, 'r', 'Baseline')],
title='All DQN Variants',
file='lunarlander_multi_all_dqn.png')