def __init__(self, experiment_parameters, run_results_dir):
self.run_results_dir = run_results_dir
self.num_tilings = check_attribute_else_default(experiment_parameters, 'num_tilings', 32)
self.tiling_length = check_attribute_else_default(experiment_parameters, 'tiling_length', 10)
self.learning_rate = check_attribute_else_default(exp_parameters, 'learning_rate', 0.001)
self.environment_name = check_attribute_else_default(experiment_parameters, 'env', 'mountain_car',
choices=['mountain_car', 'catcher'])
self.verbose = experiment_parameters.verbose
self.config = Config()
self.config.store_summary = True
self.summary = {}
""" Parameters for the Environment """
self.config.max_actions = ENVIRONMENT_DICTIONARY[self.environment_name]['max_actions']
self.config.norm_state = True
""" Parameters for the Function Approximator """
self.config.state_dims = ENVIRONMENT_DICTIONARY[self.environment_name]['state_dims']
self.config.num_actions = ENVIRONMENT_DICTIONARY[self.environment_name]['num_actions']
self.config.gamma = 1.0
self.config.epsilon = 0.1
self.config.lr = self.learning_rate / self.num_tilings
self.config.num_tilings = self.num_tilings
self.config.tiling_length = self.tiling_length
self.config.scaling_factor = 1/2
self.config.scaling_offset = 1
self.env = ENVIRONMENT_DICTIONARY[self.environment_name]['class'](config=self.config, summary=self.summary)
self.fa = TileCoderFA(config=self.config)
self.rl_agent = Agent(environment=self.env, function_approximator=self.fa, config=self.config,
summary=self.summary)
def __init__(self, experiment_parameters, run_results_dir):
self.run_results_dir = run_results_dir
self.tnet_update_Freq = check_attribute(experiment_parameters,
'tnet_update_freq', 1)
self.buffer_size = check_attribute(experiment_parameters,
'buffer_size', 10000)
self.learning_rate = check_attribute(exp_parameters, 'lr', 0.001)
self.environment_name = check_attribute(
experiment_parameters,
'env',
'mountain_car',
choices=['mountain_car', 'catcher', 'puddle_world'])
self.ppa = check_attribute(exp_parameters, 'ppa', 0.1)
self.gated = check_attribute(exp_parameters, 'gated', False)
self.gate_function = check_attribute(exp_parameters, 'gate_function',
'tanh')
self.verbose = experiment_parameters.verbose
self.config = Config()
self.config.store_summary = True
# stored in summary: 'return_per_episode', 'loss_per_step', 'steps_per_episode', 'reward_per_step'
self.summary = {}
self.config.number_of_steps = ENVIRONMENT_DICTIONARY[
self.environment_name]['number_of_steps']
""" Parameters for the Environment """
self.config.max_episode_length = ENVIRONMENT_DICTIONARY[
self.environment_name]['max_episode_length']
self.config.norm_state = True
self.config.current_step = 0
""" Parameters for the Function Approximator """
self.config.state_dims = ENVIRONMENT_DICTIONARY[
self.environment_name]['state_dims']
self.config.num_actions = ENVIRONMENT_DICTIONARY[
self.environment_name]['num_actions']
self.config.gamma = 1.0
self.config.epsilon = 0.1
self.config.optim = "adam"
self.config.lr = self.learning_rate
self.config.batch_size = 32
self.config.ppa = self.ppa
self.config.h1_dims = 32
self.config.h2_dims = 256
self.config.gate_function = self.gate_function
# DQN parameters
self.config.buffer_size = self.buffer_size
self.config.tnet_update_freq = self.tnet_update_Freq
self.config.input_dims = self.config.state_dims
if self.gated:
self.config.gated = True
else:
self.config.gated = False
self.env = ENVIRONMENT_DICTIONARY[self.environment_name]['class'](
config=self.config, summary=self.summary)
self.fa = ActionDQN(config=self.config, summary=self.summary)
self.rl_agent = Agent(environment=self.env,
function_approximator=self.fa,
config=self.config,
summary=self.summary)
def __init__(self, experiment_parameters, run_results_dir):
self.run_results_dir = run_results_dir
self.buffer_size = check_attribute_else_default(experiment_parameters, 'buffer_size', 20000)
self.tnet_update_freq = check_attribute_else_default(experiment_parameters, 'tnet_update_freq', 10)
self.environment_name = check_attribute_else_default(experiment_parameters, 'env', 'mountain_car',
choices=['mountain_car', 'catcher'])
self.verbose = experiment_parameters.verbose
# parameters specific to the parameter sweep
self.learning_rate = check_attribute_else_default(exp_parameters, 'lr', 0.001)
self.l1_reg = check_attribute_else_default(experiment_parameters, 'l1_reg', True)
self.weights_reg = check_attribute_else_default(experiment_parameters, 'weights_reg', True)
self.reg_factor = check_attribute_else_default(experiment_parameters, 'reg_factor', 0.1)
self.config = Config()
self.config.store_summary = True
# stored in summary: 'return_per_episode', 'loss_per_step', 'steps_per_episode', 'reward_per_step'
self.summary = {}
self.config.number_of_steps = ENVIRONMENT_DICTIONARY[self.environment_name]['number_of_steps']
""" Parameters for the Environment """
# Same for every experiment
self.config.max_episode_length = ENVIRONMENT_DICTIONARY[self.environment_name]['max_episode_length']
self.config.norm_state = True
self.config.current_step = 0
""" Parameters for the Function Approximator """
# Same for every experiment
self.config.state_dims = ENVIRONMENT_DICTIONARY[self.environment_name]['state_dims']
self.config.num_actions = ENVIRONMENT_DICTIONARY[self.environment_name]['num_actions']
self.config.gamma = 1.0
self.config.epsilon = 0.1
self.config.optim = "adam"
self.config.batch_size = 32
# Selected after finding the best parameter combinations for DQN with a given buffer size
self.config.buffer_size = self.buffer_size
self.config.tnet_update_freq = self.tnet_update_freq
# These are the parameters that we are sweeping over
self.config.lr = self.learning_rate
self.config.reg_method = 'l1' if self.l1_reg else 'l2'
self.config.weights_reg = self.weights_reg
self.config.reg_factor = self.reg_factor
self.env = ENVIRONMENT_DICTIONARY[self.environment_name]['class'](config=self.config, summary=self.summary)
self.fa = RegularizedNeuralNetwork(config=self.config, summary=self.summary)
self.rl_agent = Agent(environment=self.env, function_approximator=self.fa, config=self.config,
summary=self.summary)
class Experiment:
def __init__(self, experiment_parameters, run_results_dir):
self.run_results_dir = run_results_dir
self.num_tilings = check_attribute_else_default(experiment_parameters, 'num_tilings', 32)
self.tiling_length = check_attribute_else_default(experiment_parameters, 'tiling_length', 10)
self.learning_rate = check_attribute_else_default(exp_parameters, 'learning_rate', 0.001)
self.environment_name = check_attribute_else_default(experiment_parameters, 'env', 'mountain_car',
choices=['mountain_car', 'catcher'])
self.verbose = experiment_parameters.verbose
self.config = Config()
self.config.store_summary = True
self.summary = {}
""" Parameters for the Environment """
self.config.max_actions = ENVIRONMENT_DICTIONARY[self.environment_name]['max_actions']
self.config.norm_state = True
""" Parameters for the Function Approximator """
self.config.state_dims = ENVIRONMENT_DICTIONARY[self.environment_name]['state_dims']
self.config.num_actions = ENVIRONMENT_DICTIONARY[self.environment_name]['num_actions']
self.config.gamma = 1.0
self.config.epsilon = 0.1
self.config.lr = self.learning_rate / self.num_tilings
self.config.num_tilings = self.num_tilings
self.config.tiling_length = self.tiling_length
self.config.scaling_factor = 1/2
self.config.scaling_offset = 1
self.env = ENVIRONMENT_DICTIONARY[self.environment_name]['class'](config=self.config, summary=self.summary)
self.fa = TileCoderFA(config=self.config)
self.rl_agent = Agent(environment=self.env, function_approximator=self.fa, config=self.config,
summary=self.summary)
def run(self):
for i in range(ENVIRONMENT_DICTIONARY[self.environment_name]['number_of_episodes']):
episode_number = i + 1
self.rl_agent.train(1)
if self.verbose and (((i+1) % 10 == 0) or i == 0):
print("Episode Number:", episode_number)
print('\tThe cumulative reward was:', self.summary['return_per_episode'][-1])
if self.environment_name == 'catcher':
assert isinstance(self.env, Catcher3)
if self.env.timeout: break
self.save_run_summary()
# self.save_tilecoder()
def save_tilecoder(self):
tilecoder_path = os.path.join(self.run_results_dir, 'tilecoder.p')
with open(tilecoder_path, mode='wb') as tilecoder_file:
pickle.dump(self.rl_agent.fa, tilecoder_file)
def save_run_summary(self):
total_reward = np.sum(self.summary['reward_per_step'])
tr_path = os.path.join(self.run_results_dir, 'total_reward.p')
with open(tr_path, mode='wb') as tr_file:
pickle.dump(total_reward, tr_file)
config_path = os.path.join(self.run_results_dir, 'config.p')
with open(config_path, mode='wb') as config_file:
pickle.dump(self.config, config_file)
class Experiment:
def __init__(self, experiment_parameters, run_results_dir):
self.run_results_dir = run_results_dir
self.tnet_update_Freq = check_attribute_else_default(
experiment_parameters, 'tnet_update_freq', 1)
self.buffer_size = check_attribute_else_default(
experiment_parameters, 'buffer_size', 10000)
self.learning_rate = check_attribute_else_default(
exp_parameters, 'lr', 0.001)
self.environment_name = check_attribute_else_default(
experiment_parameters,
'env',
'mountain_car',
choices=['mountain_car', 'catcher', 'puddle_world'])
self.verbose = experiment_parameters.verbose
self.config = Config()
self.config.store_summary = True
# stored in summary: 'return_per_episode', 'loss_per_step', 'steps_per_episode', 'reward_per_step'
self.summary = {}
self.config.number_of_steps = ENVIRONMENT_DICTIONARY[
self.environment_name]['number_of_steps']
""" Parameters for the Environment """
self.config.max_episode_length = ENVIRONMENT_DICTIONARY[
self.environment_name]['max_episode_length']
self.config.norm_state = True
self.config.current_step = 0
""" Parameters for the Function Approximator """
self.config.state_dims = ENVIRONMENT_DICTIONARY[
self.environment_name]['state_dims']
self.config.num_actions = ENVIRONMENT_DICTIONARY[
self.environment_name]['num_actions']
self.config.gamma = 1.0
self.config.epsilon = 0.1
self.config.optim = "adam"
self.config.lr = self.learning_rate
self.config.batch_size = 32
# DQN parameters
self.config.buffer_size = self.buffer_size
self.config.tnet_update_freq = self.tnet_update_Freq
self.env = ENVIRONMENT_DICTIONARY[self.environment_name]['class'](
config=self.config, summary=self.summary)
self.fa = VanillaDQN(config=self.config, summary=self.summary)
self.rl_agent = Agent(environment=self.env,
function_approximator=self.fa,
config=self.config,
summary=self.summary)
def run(self):
prev_idx = 0
current_episode_number = 1
while self.config.current_step != self.config.number_of_steps:
self.rl_agent.train(1)
if self.verbose and ((current_episode_number % 10 == 0) or
(current_episode_number - 1 == 0)):
print("Episode Number:", current_episode_number)
print('\tThe cumulative reward was:',
self.summary['return_per_episode'][-1])
print(
'\tThe cumulative loss was:',
np.round(np.sum(self.summary['loss_per_step'][prev_idx:]),
2))
print('\tCurrent environment steps:', self.config.current_step)
prev_idx = self.config.current_step
current_episode_number += 1
if self.verbose:
print("Number of episodes completed:",
len(self.summary['return_per_episode']))
print("The total cumulative reward was:",
np.sum(self.summary['reward_per_step']))
print("Current environment steps:", self.config.current_step)
self.save_network_params()
self.save_run_summary()
def save_network_params(self):
params_path = os.path.join(self.run_results_dir,
'final_network_weights.pt')
torch.save(self.fa.net.state_dict(), params_path)
def save_run_summary(self):
summary_path = os.path.join(self.run_results_dir, 'summary.p')
with open(summary_path, mode='wb') as summary_file:
pickle.dump(self.summary, summary_file)
config_path = os.path.join(self.run_results_dir, 'config.p')
with open(config_path, mode='wb') as config_file:
pickle.dump(self.config, config_file)
def __init__(self, experiment_parameters, run_results_dir):
self.run_results_dir = run_results_dir
self.learning_rate = check_attribute_else_default(
exp_parameters, 'lr', 0.001)
self.buffer_size = check_attribute_else_default(
experiment_parameters, 'buffer_size', 20000)
self.tnet_update_freq = check_attribute_else_default(
experiment_parameters, 'tnet_update_freq', 10)
self.environment_name = check_attribute_else_default(
experiment_parameters,
'env',
'mountain_car',
choices=['mountain_car', 'catcher'])
self.verbose = experiment_parameters.verbose
# parameters specific to distributional regularizers
self.beta = check_attribute_else_default(experiment_parameters, 'beta',
0.1)
self.reg_factor = check_attribute_else_default(experiment_parameters,
'reg_factor', 0.1)
self.use_gamma = check_attribute_else_default(experiment_parameters,
'use_gamma', False)
self.beta_lb = check_attribute_else_default(experiment_parameters,
'beta_lb', False)
self.config = Config()
self.config.store_summary = True
# stored in summary: 'return_per_episode', 'loss_per_step', 'steps_per_episode', 'reward_per_step'
self.summary = {}
self.config.number_of_steps = ENVIRONMENT_DICTIONARY[
self.environment_name]['number_of_steps']
""" Parameters for the Environment """
self.config.max_episode_length = ENVIRONMENT_DICTIONARY[
self.environment_name]['max_episode_length']
self.config.norm_state = True
self.config.current_step = 0
""" Parameters for the Function Approximator """
self.config.state_dims = ENVIRONMENT_DICTIONARY[
self.environment_name]['state_dims']
self.config.num_actions = ENVIRONMENT_DICTIONARY[
self.environment_name]['num_actions']
self.config.gamma = 1.0
self.config.epsilon = 0.1
self.config.optim = "adam"
self.config.batch_size = 32
# DQN parameters
self.config.lr = self.learning_rate
self.config.buffer_size = self.buffer_size
self.config.tnet_update_freq = self.tnet_update_freq
self.config.beta = self.beta
self.config.reg_factor = self.reg_factor
self.config.use_gamma = self.use_gamma
self.config.beta_lb = self.beta_lb
self.env = ENVIRONMENT_DICTIONARY[self.environment_name]['class'](
config=self.config, summary=self.summary)
self.fa = DistRegNeuralNetwork(config=self.config,
summary=self.summary)
self.rl_agent = Agent(environment=self.env,
function_approximator=self.fa,
config=self.config,
summary=self.summary)
def __init__(self, experiment_parameters, run_results_dir):
self.run_results_dir = run_results_dir
self.buffer_size = check_attribute_else_default(
experiment_parameters, 'buffer_size', 20000)
self.method = check_attribute_else_default(exp_parameters, 'method',
'DQN')
self.environment_name = check_attribute_else_default(
experiment_parameters,
'env',
'mountain_car',
choices=['mountain_car', 'catcher', 'puddle_world'])
parameters_dictionary = BEST_PARAMETERS_DICTIONARY[
self.environment_name][self.method][self.buffer_size]
self.verbose = experiment_parameters.verbose
self.config = Config()
self.config.store_summary = True
# stored in summary: 'return_per_episode', 'loss_per_step', 'steps_per_episode', 'reward_per_step'
self.summary = {}
self.config.number_of_steps = ENVIRONMENT_DICTIONARY[
self.environment_name]['number_of_steps']
""" Parameters for the Environment """
self.config.max_episode_length = ENVIRONMENT_DICTIONARY[
self.environment_name]['max_episode_length']
self.config.norm_state = True
self.config.current_step = 0
""" Parameters for the Function Approximator """
self.config.state_dims = ENVIRONMENT_DICTIONARY[
self.environment_name]['state_dims']
self.config.num_actions = ENVIRONMENT_DICTIONARY[
self.environment_name]['num_actions']
self.config.gamma = 1.0
self.config.epsilon = 0.1
self.config.optim = "adam"
self.config.batch_size = 32
# Parameters for any type of agent
self.config.buffer_size = self.buffer_size
self.config.lr = parameters_dictionary['LearningRate']
self.config.tnet_update_freq = parameters_dictionary['Freq']
if self.method in ['DRE', 'DRE_LB', 'DRG', 'DRG_LB']:
self.config.beta = parameters_dictionary['Beta']
self.config.reg_factor = parameters_dictionary['RegFactor']
self.config.use_gamma = False
self.config.beta_lb = False
if self.method in ['DRG', 'DRG_LB']:
self.config.use_gamma = True
if self.method in ['DRE_LB', 'DRG_LB']:
self.config.beta_lb = True
self.fa = DistRegNeuralNetwork(config=self.config,
summary=self.summary)
elif self.method in ['L1A', 'L1W', 'L2A', 'L2W']:
self.config.reg_factor = parameters_dictionary['RegFactor']
self.config.reg_method = 'l1'
if self.method in ['L2A', 'L2W']:
self.config.reg_method = 'l2'
self.config.weights_reg = False
if self.method in ['L1W', 'L2W']:
self.config.weights_reg = True
self.fa = RegularizedNeuralNetwork(config=self.config,
summary=self.summary)
elif self.method in ['DQN']:
self.fa = VanillaDQN(config=self.config, summary=self.summary)
elif self.method in ['Dropout']:
self.config.dropout_probability = parameters_dictionary[
'DropoutProbability']
self.fa = DropoutNeuralNetwork(config=self.config,
summary=self.summary)
else:
raise ValueError(
"No configuration available for the given method.")
self.env = ENVIRONMENT_DICTIONARY[self.environment_name]['class'](
config=self.config, summary=self.summary)
self.rl_agent = Agent(environment=self.env,
function_approximator=self.fa,
config=self.config,
summary=self.summary)
class Experiment:
def __init__(self, experiment_parameters, run_results_dir):
self.run_results_dir = run_results_dir
self.buffer_size = check_attribute_else_default(
experiment_parameters, 'buffer_size', 20000)
self.method = check_attribute_else_default(exp_parameters, 'method',
'DQN')
self.environment_name = check_attribute_else_default(
experiment_parameters,
'env',
'mountain_car',
choices=['mountain_car', 'catcher', 'puddle_world'])
parameters_dictionary = BEST_PARAMETERS_DICTIONARY[
self.environment_name][self.method][self.buffer_size]
self.verbose = experiment_parameters.verbose
self.config = Config()
self.config.store_summary = True
# stored in summary: 'return_per_episode', 'loss_per_step', 'steps_per_episode', 'reward_per_step'
self.summary = {}
self.config.number_of_steps = ENVIRONMENT_DICTIONARY[
self.environment_name]['number_of_steps']
""" Parameters for the Environment """
self.config.max_episode_length = ENVIRONMENT_DICTIONARY[
self.environment_name]['max_episode_length']
self.config.norm_state = True
self.config.current_step = 0
""" Parameters for the Function Approximator """
self.config.state_dims = ENVIRONMENT_DICTIONARY[
self.environment_name]['state_dims']
self.config.num_actions = ENVIRONMENT_DICTIONARY[
self.environment_name]['num_actions']
self.config.gamma = 1.0
self.config.epsilon = 0.1
self.config.optim = "adam"
self.config.batch_size = 32
# Parameters for any type of agent
self.config.buffer_size = self.buffer_size
self.config.lr = parameters_dictionary['LearningRate']
self.config.tnet_update_freq = parameters_dictionary['Freq']
if self.method in ['DRE', 'DRE_LB', 'DRG', 'DRG_LB']:
self.config.beta = parameters_dictionary['Beta']
self.config.reg_factor = parameters_dictionary['RegFactor']
self.config.use_gamma = False
self.config.beta_lb = False
if self.method in ['DRG', 'DRG_LB']:
self.config.use_gamma = True
if self.method in ['DRE_LB', 'DRG_LB']:
self.config.beta_lb = True
self.fa = DistRegNeuralNetwork(config=self.config,
summary=self.summary)
elif self.method in ['L1A', 'L1W', 'L2A', 'L2W']:
self.config.reg_factor = parameters_dictionary['RegFactor']
self.config.reg_method = 'l1'
if self.method in ['L2A', 'L2W']:
self.config.reg_method = 'l2'
self.config.weights_reg = False
if self.method in ['L1W', 'L2W']:
self.config.weights_reg = True
self.fa = RegularizedNeuralNetwork(config=self.config,
summary=self.summary)
elif self.method in ['DQN']:
self.fa = VanillaDQN(config=self.config, summary=self.summary)
elif self.method in ['Dropout']:
self.config.dropout_probability = parameters_dictionary[
'DropoutProbability']
self.fa = DropoutNeuralNetwork(config=self.config,
summary=self.summary)
else:
raise ValueError(
"No configuration available for the given method.")
self.env = ENVIRONMENT_DICTIONARY[self.environment_name]['class'](
config=self.config, summary=self.summary)
self.rl_agent = Agent(environment=self.env,
function_approximator=self.fa,
config=self.config,
summary=self.summary)
def run(self):
prev_idx = 0
current_episode_number = 1
assert hasattr(self.config, 'current_step')
while self.config.current_step != self.config.number_of_steps:
self.rl_agent.train(1)
if self.verbose and ((current_episode_number % 10 == 0) or
(current_episode_number - 1 == 0)):
print("Episode Number:", current_episode_number)
print('\tThe cumulative reward was:',
self.summary['return_per_episode'][-1])
print(
'\tThe cumulative loss was:',
np.round(np.sum(self.summary['loss_per_step'][prev_idx:]),
2))
print('\tCurrent environment steps:', self.config.current_step)
prev_idx = self.config.current_step
current_episode_number += 1
if self.verbose:
print("Number of episodes completed:",
len(self.summary['return_per_episode']))
print("The total cumulative reward was:",
np.sum(self.summary['reward_per_step']))
print("Current environment steps:", self.config.current_step)
self.save_network_params()
self.save_run_summary()
def save_network_params(self):
params_path = os.path.join(self.run_results_dir,
'final_network_weights.pt')
torch.save(self.fa.net.state_dict(), params_path)
def save_run_summary(self):
summary_path = os.path.join(self.run_results_dir, 'summary.p')
with open(summary_path, mode='wb') as summary_file:
pickle.dump(self.summary, summary_file)
config_path = os.path.join(self.run_results_dir, 'config.p')
with open(config_path, mode='wb') as config_file:
pickle.dump(self.config, config_file)