def main():
user_input_module = user_input.UserInputModule(is_asyn=False)
env = gym.make('Assault-v0')
for i_episode in range(20):
observation = env.reset()
for t in range(100):
env.render()
print(observation)
human_action = user_input_module.getInput()
if human_action == 'j':
action = 0
elif human_action == 'f':
action = 1
elif human_action == 'd':
action = 2
elif human_action == 's':
action = 3
elif human_action == 'a':
action = 4
elif human_action == 'k':
action = 5
else:
action = 0
print(action)
observation, reward, done, info = env.step(action)
if done:
print("Episode finished after {} time steps".format(t + 1))
break
env.close()
def run_episodes(self):
user_input_module = user_input.UserInputModule(is_asyn=True)
env = gym.make('Assault-v0')
human_actions = {}
for i_episode in range(5):
episode_log_dir = self.expr_log_dir + '/' + 'episode-' + str(
i_episode)
if not os.path.exists(episode_log_dir):
os.makedirs(episode_log_dir)
env.reset()
for t in range(10000):
env.render()
env.env.ale.saveScreenPNG(
six.b(episode_log_dir + '/' + str(t) + '.png'))
human_action = user_input_module.getInput()
if human_action is None:
action = env.action_space.sample()
else:
print(human_action)
human_actions[str(i_episode) + '_' + str(t)] = human_action
if human_action == 'j':
action = 0
elif human_action == 'f':
action = 1
elif human_action == 'd':
action = 2
elif human_action == 's':
action = 3
elif human_action == 'a':
action = 4
elif human_action == 'k':
action = 5
else:
action = 0
observation, reward, done, info = env.step(action)
if done:
print("Episode finished after {} time steps".format(t + 1))
break
counter_saver.saveDictToFile(
self.expr_log_dir + '/' + 'human_actions.txt', human_actions)
env.close()
def __init__(self, grid_name='DiscountGrid', discount=0.9, learning_rate=0.5, living_reward=0.0
, noise=0, epsilon=0.3, display_speed=0.5
, grid_size=150, text_only=False, n_episodes=100
, agent_window_size=1
, agent_max_n_experiences=1000
, check_policy_converge=False
, optimal_policy=None
, expr_log_dir=None
, agent_type="qLearningAgent"
, init_temp=1024.0
, temp_decrease_rate=2.0
, is_asyn_input=True):
"""
:param agent_type: "qLearningAgent" or "TamerAgent" or "preferenceTAMERAgent"
"""
###########################
# GENERAL CONTROL
###########################
self.text_only = text_only
self.display_speed = display_speed
self.n_episodes = n_episodes
self.discount = discount
self.check_policy_converge = check_policy_converge
self.optimal_policy = optimal_policy
self.expr_log_dir = expr_log_dir
self.save_VDBE = False
###########################
# GET THE INPUT MODULE
###########################
if agent_type == qlearningAgents.QLearningAgent.getAgentType():
self.user_input_module = None
else:
self.user_input_module = user_input.UserInputModule(is_asyn=is_asyn_input)
###########################
# GET THE GRIDWORLD
###########################
# noinspection PyUnresolvedReferences
import gridworld
mdp_function = getattr(gridworld, "get" + grid_name)
self.mdp = mdp_function()
self.mdp.setLivingReward(living_reward)
self.mdp.setNoise(noise)
self.env = gridworld.GridworldEnvironment(self.mdp)
###########################
# Variables used to store parameters values
###########################
# init VDBE values records
global VDBE_RECORDS
VDBE_RECORDS = dict()
for state in self.env.getGridWorld().getNonTerminalStates():
VDBE_RECORDS[state] = list()
###########################
# GET THE DISPLAY ADAPTER
###########################
import textGridworldDisplay
self.display = textGridworldDisplay.TextGridworldDisplay(self.mdp)
if not text_only:
import graphicsGridworldDisplay
self.display = graphicsGridworldDisplay.GraphicsGridworldDisplay(self.mdp, grid_size, display_speed)
try:
self.display.start()
except KeyboardInterrupt:
sys.exit(0)
###########################
# GET THE TAMER AGENT
###########################
# env.getPossibleActions, opts.discount, opts.learningRate, opts.epsilon
# simulationFn = lambda agent, state: simulation.GridworldSimulation(agent,state,mdp)
self.gridWorldEnv = GridworldEnvironment(self.mdp)
action_function = lambda m_state: self.mdp.getPossibleActions(m_state)
q_learn_opts = {
'gamma': discount,
'alpha': learning_rate,
'epsilon': epsilon,
'actionFn': action_function,
'init_temp': init_temp,
'temp_decrease_rate': temp_decrease_rate
}
if agent_type == qlearningAgents.QLearningAgent.getAgentType():
self.agent = qlearningAgents.QLearningAgent(**q_learn_opts)
elif agent_type == qlearningAgents.TamerQAgent.getAgentType():
self.agent = qlearningAgents.TamerQAgent(max_n_experiences=agent_max_n_experiences
, window_size=agent_window_size
, is_asyn_input=is_asyn_input
, **q_learn_opts)
elif agent_type == preferenceTamerAgent.PreferenceTAMERAgent.getAgentType():
self.agent = preferenceTamerAgent.PreferenceTAMERAgent(max_n_experiences=agent_max_n_experiences
, window_size=agent_window_size
, is_asyn_input=is_asyn_input
, **q_learn_opts)
def __init__(self,
grid_name='DiscountGrid',
discount=0.9,
learning_rate=0.5,
living_reward=0.0,
noise=0,
epsilon=0.3,
display_speed=0.5,
grid_size=150,
text_only=False,
n_episodes=100,
agent_window_size=1,
agent_max_n_experiences=1000,
check_value_converge=False,
check_policy_converge=False,
optimal_policy=None,
expr_log_dir=None,
delta=0.02,
is_use_q_agent=False,
init_temp=1024.0,
temp_decrease_rate=2.0,
is_asyn_input=True):
###########################
# GENERAL CONTROL
###########################
self.text_only = text_only
self.display_speed = display_speed
self.n_episodes = n_episodes
self.discount = discount
self.check_value_converge = check_value_converge
self.check_policy_converge = check_policy_converge
self.optimal_policy = optimal_policy
self.expr_log_dir = expr_log_dir
self.delta = delta
###########################
# GET THE INPUT MODULE
###########################
if is_use_q_agent:
self.user_input_module = None
else:
self.user_input_module = user_input.UserInputModule(
is_asyn=is_asyn_input)
###########################
# GET THE GRIDWORLD
###########################
# noinspection PyUnresolvedReferences
import gridworld
mdp_function = getattr(gridworld, "get" + grid_name)
self.mdp = mdp_function()
self.mdp.setLivingReward(living_reward)
self.mdp.setNoise(noise)
self.env = gridworld.GridworldEnvironment(self.mdp)
###########################
# GET THE DISPLAY ADAPTER
###########################
import textGridworldDisplay
self.display = textGridworldDisplay.TextGridworldDisplay(self.mdp)
if not text_only:
import graphicsGridworldDisplay
self.display = graphicsGridworldDisplay.GraphicsGridworldDisplay(
self.mdp, grid_size, display_speed)
try:
self.display.start()
except KeyboardInterrupt:
sys.exit(0)
###########################
# GET THE TAMER AGENT
###########################
import qlearningAgents
# env.getPossibleActions, opts.discount, opts.learningRate, opts.epsilon
# simulationFn = lambda agent, state: simulation.GridworldSimulation(agent,state,mdp)
self.gridWorldEnv = GridworldEnvironment(self.mdp)
action_function = lambda state: self.mdp.getPossibleActions(state)
q_learn_opts = {
'gamma': discount,
'alpha': learning_rate,
'epsilon': epsilon,
'actionFn': action_function,
'init_temp': init_temp,
'temp_decrease_rate': temp_decrease_rate
}
if is_use_q_agent:
self.agent = qlearningAgents.QLearningAgent(**q_learn_opts)
else:
self.agent = qlearningAgents.TamerQAgent(
max_n_experiences=agent_max_n_experiences,
window_size=agent_window_size,
is_asyn_input=is_asyn_input,
**q_learn_opts)
def __init__(self, optimal_policy=None, expr_log_dir=None):
###########################
# GENERAL CONTROL
###########################
self.text_only = ExperimentConfigurator.experimentConfig['text_only']
self.display_speed = ExperimentConfigurator.experimentConfig[
'display_speed']
self.n_episodes = ExperimentConfigurator.gridWorldConfig['n_episodes']
self.discount = ExperimentConfigurator.gridWorldConfig['discount']
self.check_policy_converge = ExperimentConfigurator.experimentConfig[
'check_policy_converge']
self.optimal_policy = optimal_policy
self.expr_log_dir = expr_log_dir
self.save_VDBE = ExperimentConfigurator.experimentConfig['save_VDBE']
###########################
# GET THE INPUT MODULE
###########################
if ExperimentConfigurator.experimentConfig[
'agent_type'] == qlearningAgents.QLearningAgent.getAgentType():
self.user_input_module = None
else:
self.user_input_module = user_input.UserInputModule(
is_asyn=ExperimentConfigurator.TamerConfig['is_asyn_input'])
self.auto_feedback = AutoFeedback()
###########################
# GET THE GRIDWORLD
###########################
# noinspection PyUnresolvedReferences
import gridworld
mdp_function = getattr(
gridworld,
"get" + ExperimentConfigurator.gridWorldConfig['grid_name'])
self.mdp = mdp_function()
self.mdp.setLivingReward(
ExperimentConfigurator.gridWorldConfig['living_reward'])
self.mdp.setNoise(ExperimentConfigurator.gridWorldConfig['noise'])
self.env = gridworld.GridworldEnvironment(self.mdp)
###########################
# Variables used to store parameters values
###########################
# init VDBE values records
global VDBE_RECORDS
VDBE_RECORDS = dict()
for state in self.env.getGridWorld().getNonTerminalStates():
VDBE_RECORDS[state] = list()
###########################
# GET THE DISPLAY ADAPTER
###########################
import textGridworldDisplay
self.display = textGridworldDisplay.TextGridworldDisplay(self.mdp)
if not self.text_only:
import graphicsGridworldDisplay
self.display = graphicsGridworldDisplay.GraphicsGridworldDisplay(
self.mdp, ExperimentConfigurator.gridWorldConfig['grid_size'],
self.display_speed)
try:
self.display.start()
except KeyboardInterrupt:
sys.exit(0)
###########################
# GET THE TAMER AGENT
###########################
# env.getPossibleActions, opts.discount, opts.learningRate, opts.epsilon
# simulationFn = lambda agent, state: simulation.GridworldSimulation(agent,state,mdp)
self.gridWorldEnv = GridworldEnvironment(self.mdp)
action_function = lambda m_state: self.mdp.getPossibleActions(m_state)
q_learn_opts = {'actionFn': action_function}
if ExperimentConfigurator.experimentConfig[
'agent_type'] == qlearningAgents.QLearningAgent.getAgentType():
self.agent = qlearningAgents.QLearningAgent(**q_learn_opts)
elif ExperimentConfigurator.experimentConfig[
'agent_type'] == qlearningAgents.TamerQAgent.getAgentType():
self.agent = qlearningAgents.TamerQAgent(
max_n_experiences=ExperimentConfigurator.
TamerConfig['agent_max_n_experiences'],
window_size=ExperimentConfigurator.
TamerConfig['agent_window_size'],
is_asyn_input=ExperimentConfigurator.
TamerConfig['is_asyn_input'],
**q_learn_opts)
elif ExperimentConfigurator.experimentConfig[
'agent_type'] == preferenceTamerAgent.PreferenceTAMERAgent.getAgentType(
):
self.agent = preferenceTamerAgent.PreferenceTAMERAgent(
max_n_experiences=ExperimentConfigurator.
TamerConfig['agent_max_n_experiences'],
window_size=ExperimentConfigurator.
TamerConfig['agent_window_size'],
is_asyn_input=ExperimentConfigurator.
TamerConfig['is_asyn_input'],
**q_learn_opts)
def __init__(self,
grid_name='DiscountGrid',
discount=0.9,
learning_rate=0.5,
living_reward=0.0,
noise=0.2,
epsilon=0.3,
display_speed=0.5,
grid_size=150,
text_only=False,
n_episodes=100,
agent_window_size=1,
agent_max_n_experiences=1000,
is_use_q_agent=False):
self.text_only = text_only
self.display_speed = display_speed
self.n_episodes = n_episodes
self.discount = discount
###########################
# GET THE INPUT MODULE
###########################
if is_use_q_agent:
self.user_input_module = None
else:
self.user_input_module = user_input.UserInputModule()
###########################
# GET THE GRIDWORLD
###########################
# noinspection PyUnresolvedReferences
import gridworld
mdp_function = getattr(gridworld, "get" + grid_name)
self.mdp = mdp_function()
self.mdp.setLivingReward(living_reward)
self.mdp.setNoise(noise)
self.env = gridworld.GridworldEnvironment(self.mdp)
###########################
# GET THE DISPLAY ADAPTER
###########################
import textGridworldDisplay
self.display = textGridworldDisplay.TextGridworldDisplay(self.mdp)
if not text_only:
import graphicsGridworldDisplay
self.display = graphicsGridworldDisplay.GraphicsGridworldDisplay(
self.mdp, grid_size, display_speed)
try:
self.display.start()
except KeyboardInterrupt:
sys.exit(0)
###########################
# GET THE TAMER AGENT
###########################
import qlearningAgents
# env.getPossibleActions, opts.discount, opts.learningRate, opts.epsilon
# simulationFn = lambda agent, state: simulation.GridworldSimulation(agent,state,mdp)
self.gridWorldEnv = GridworldEnvironment(self.mdp)
action_function = lambda state: self.mdp.getPossibleActions(state)
q_learn_opts = {
'gamma': discount,
'alpha': learning_rate,
'epsilon': epsilon,
'actionFn': action_function
}
if is_use_q_agent:
self.agent = qlearningAgents.QLearningAgent(**q_learn_opts)
else:
self.agent = qlearningAgents.TamerQAgent(
max_n_experiences=agent_max_n_experiences,
window_size=agent_window_size,
**q_learn_opts)
