def __init__(self):
""" @brief: Setting up internal parameters for the RL module"""
# Navigation Task
self._environment = NavigationEnvironment()
self._task = NavigationTask(self._environment)
# Number of States : (read from params.py)
self._states = STATES
self._state_limits = LIMITS
# Total number of states:
self._number_of_states = 1
for i in self._states:
self._number_of_states *= i
# Number of actions
self._actions = ACTION_STATES
self._action_limits = ACTION_RANGE
# Action Value Table directory
self.tables_directory = os.path.dirname(__file__) + "/tables/"
self.table_code = "S"+str(self._number_of_states)+"_"+"A"+str(self._actions)
self._filename = FILENAME + self.table_code
# Action Value Table setup
self.load_AV_Table()
# Declare ROS Service to store Action Value Table
store_service = rospy.Service('store_table', StoreAVTable, self.store_cb)
# Set up task parameters:
self._task.set_params(COMMAND_DURATION,
FUSION_WEIGHTS,
TIME_GRANULARITY,
self._state_limits,
MAX_REWARD,
COST_THRESHOLD)
# Agent set up
self._learner = SARSA(alpha,gamma)
self._learner._setExplorer(EpsilonGreedyExplorer(epsilon))
self._agent = LearningAgent(self._av_table, self._learner)
# Experiment set up
self._experiment = Experiment(self._task,self._agent)
self._experiment.set_params(STEP_SIZE)
# Start print table thread
if VISUALIZATION is True:
try:
#thread.start_new_thread(self.print_table,())
self.visualization_thread = Thread(target = self.print_table, args = () )
self.visualization_thread.start()
except:
print "Failed to start visualization thread!"
print "Successfully Initialization of RL module! (kappa)"
def __init__(self): self.av_table = ActionValueTable(2, 3) self.av_table.initialize(0.1) learner = SARSA() learner._setExplorer(EpsilonGreedyExplorer(0.0)) self.agent = LearningAgent(self.av_table, learner) env = HASSHEnv() task = HASSHTask(env) self.experiment = Experiment(task, self.agent)
def __init__(self): self.av_table = ActionValueTable(4, 5) self.av_table.initialize(0.1) learner = SARSA() learner._setExplorer(EpsilonGreedyExplorer(0.0)) self.agent = LearningAgent(self.av_table, learner) env = HASSHEnv() task = HASSHTask(env) self.experiment = Experiment(task, self.agent)
def setup_RL():
# create the maze with walls (1)
envmatrix = np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1],
[1, 0, 0, 1, 0, 0, 0, 0, 1],
[1, 0, 0, 1, 0, 0, 1, 0, 1],
[1, 0, 0, 1, 0, 0, 1, 0, 1],
[1, 0, 0, 1, 0, 1, 1, 0, 1],
[1, 0, 0, 0, 0, 0, 1, 0, 1],
[1, 1, 1, 1, 1, 1, 1, 0, 1],
[1, 0, 0, 0, 0, 0, 0, 0, 1],
[1, 1, 1, 1, 1, 1, 1, 1, 1]])
env = Maze(envmatrix, (7, 7))
# create task
task = MDPMazeTask(env)
# create value table and initialize with ones
table = ActionValueTable(81, 4)
table.initialize(0.)
# create agent with controller and learner - use SARSA(), Q() or QLambda() here
# learner = Q()
learner = SARSA()
# create agent
agent = LearningAgent(table, learner)
# create experiment
experiment = Experiment(task, agent)
return experiment, agent, table
def createAgent(module):
# create agent with controller and learner - use SARSA(), Q() or QLambda() here
## alpha -- learning rate (preference of new information)
## gamma -- discount factor (importance of future reward)
# learner = Q(0.5, 0.99)
learner = SARSA(0.5, 0.99)
# learner = QLambda(0.5, 0.99, 0.9)
agent = LearningAgent(module, learner)
return agent
def createAgent(module):
### create agent with controller and learner - use SARSA(), Q() or QLambda() here
## alpha -- learning rate (preference of new information -- update value factor)
## gamma -- discount factor (importance of future reward -- next value factor)
learner = SARSA(alpha=0.3, gamma=0.99)
# learner = Q(alpha=0.3, gamma=0.99)
explorer = learner.explorer
explorer.epsilon = 0.4
explorer.decay = 0.9999
agent = LearningAgent(module, learner)
return agent
def testValueBased(self):
""" Test value-based learner.
"""
mkt = SmartMarket(self.case)
exp = MarketExperiment([], [], mkt)
for g in self.case.generators:
env = DiscreteMarketEnvironment([g], mkt)
dim_state, num_actions = (10, 10)
exp.tasks.append(ProfitTask(env, dim_state, num_actions))
module = ActionValueTable(dim_state, num_actions)
module.initialize(1.0)
# module = ActionValueNetwork(dimState=1, numActions=4)
learner = SARSA() #Q() QLambda()
# learner.explorer = BoltzmannExplorer() # default is e-greedy.
exp.agents.append(LearningAgent(module, learner))
for _ in range(1000):
exp.doInteractions(24) # interact with the env in batch mode
for agent in exp.agents:
agent.learn()
agent.reset()
warnings.filterwarnings("ignore")
# create the maze with walls (1)
envmatrix = array([[1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 0, 0, 1, 0, 0, 0, 0, 1],
[1, 0, 0, 1, 0, 0, 1, 0, 1], [1, 0, 0, 1, 0, 0, 1, 0, 1],
[1, 0, 0, 1, 0, 1, 1, 0, 1], [1, 0, 0, 0, 0, 0, 1, 0, 1],
[1, 1, 1, 1, 1, 1, 1, 0, 1], [1, 0, 0, 0, 0, 0, 0, 0, 1],
[1, 1, 1, 1, 1, 1, 1, 1, 1]])
env = Maze(envmatrix, (7, 7))
# create task
task = MDPMazeTask(env)
# create value table and initialize with ones
table = ActionValueTable(81, 4)
table.initialize(1.)
# create agent with controller and learner - use SARSA(), Q() or QLambda() here
# learner = Q()
learner = SARSA()
# standard exploration is e-greedy, but a different type can be chosen as well
# learner.explorer = BoltzmannExplorer()
# create agent
agent = LearningAgent(table, learner)
# create experiment
# experiment = Experiment(task, agent)
experiment = EpisodicExperiment(task, agent)
# prepare plotting
pylab.gray()
pylab.ion()
for i in range(50):
# interact with the environment (here in batch mode)
experiment.doInteractions(100)
agent.learn()
agent.reset()
task = GymTask.createTask(gymRawEnv) env = task.env env.setTransformation(transformation) env.setCumulativeRewardMode() # create value table and initialize with ones table = ActionValueTable(observationDigitizer.states, env.numActions) table.initialize(0.0) # table.initialize( np.random.rand( table.paramdim ) ) # create agent with controller and learner - use SARSA(), Q() or QLambda() here ## alpha -- learning rate (preference of new information) ## gamma -- discount factor (importance of future reward) # learner = Q(0.5, 0.99) learner = SARSA(0.5, 0.99) # learner = QLambda(0.5, 0.99, 0.9) explorer = learner.explorer explorer.decay = 0.999992 agent = LearningAgent(table, learner) experiment = Experiment(task, agent) ## prevents "ImportError: sys.meta_path is None, Python is likely shutting down" atexit.register(task.close) render_demo = False render_steps = False imax = 7000 period_print = 100
gymRawEnv = gym.make('FrozenLake-v0')
task = GymTask.createTask(gymRawEnv)
task.env.setTransformation(EnvTransformation())
# create value table and initialize with ones
table = ActionValueTable(gymRawEnv.observation_space.n,
gymRawEnv.action_space.n)
table.initialize(0.0)
# table.initialize( np.random.rand( table.paramdim ) )
### create agent with controller and learner - use SARSA(), Q() or QLambda() here
## alpha -- learning rate (preference of new information -- update value factor)
## gamma -- discount factor (importance of future reward -- next value factor)
learner = SARSA(alpha=0.3, gamma=0.99)
# learner = Q(alpha=0.3, gamma=0.99)
explorer = learner.explorer
explorer.epsilon = 0.4
explorer.decay = 0.9999
agent = LearningAgent(table, learner)
experiment = Experiment(task, agent)
render_steps = False
imax = 5000
# prepare plotting
if render_steps:
pylab.gray()
class KinodynamicController(object):
""" @brief: A class to handle interactions between various components of the RL module"""
def __init__(self):
""" @brief: Setting up internal parameters for the RL module"""
# Navigation Task
self._environment = NavigationEnvironment()
self._task = NavigationTask(self._environment)
# Number of States : (read from params.py)
self._states = STATES
self._state_limits = LIMITS
# Total number of states:
self._number_of_states = 1
for i in self._states:
self._number_of_states *= i
# Number of actions
self._actions = ACTION_STATES
self._action_limits = ACTION_RANGE
# Action Value Table directory
self.tables_directory = os.path.dirname(__file__) + "/tables/"
self.table_code = "S"+str(self._number_of_states)+"_"+"A"+str(self._actions)
self._filename = FILENAME + self.table_code
# Action Value Table setup
self.load_AV_Table()
# Declare ROS Service to store Action Value Table
store_service = rospy.Service('store_table', StoreAVTable, self.store_cb)
# Set up task parameters:
self._task.set_params(COMMAND_DURATION,
FUSION_WEIGHTS,
TIME_GRANULARITY,
self._state_limits,
MAX_REWARD,
COST_THRESHOLD)
# Agent set up
self._learner = SARSA(alpha,gamma)
self._learner._setExplorer(EpsilonGreedyExplorer(epsilon))
self._agent = LearningAgent(self._av_table, self._learner)
# Experiment set up
self._experiment = Experiment(self._task,self._agent)
self._experiment.set_params(STEP_SIZE)
# Start print table thread
if VISUALIZATION is True:
try:
#thread.start_new_thread(self.print_table,())
self.visualization_thread = Thread(target = self.print_table, args = () )
self.visualization_thread.start()
except:
print "Failed to start visualization thread!"
print "Successfully Initialization of RL module! (kappa)"
def store_cb(self,req):
storeData(self._av_table,self._filename)
print "Saved AV Table"
return True
def load_AV_Table(self):
load_D = loadData(self._filename)
if load_D[1] == True:
self._av_table = load_D[0]
print "Found Table!"
else:
self._av_table = ActionValueTable(self._number_of_states, self._actions)
self._av_table.initialize(0.0)
print "No training for this format. Creating new AV table"
def print_table(self):
""" @brief: Visual Representation of Action Value Table
@return: nothing
"""
matplotlib.pyplot.ion()
while True:
data = self._av_table.params.reshape(self._number_of_states,self._actions)
matplotlib.pyplot.pcolor(data,
cmap = matplotlib.pyplot.cm.RdYlGn,
vmin = -MAX_REWARD,
vmax = MAX_REWARD)
matplotlib.pyplot.draw()
rospy.sleep(2)
def __del__(self):
# Terminate visualization thread
if VISUALIZATION is True:
self.visualization_thread.join()
# Copy learned data to repo
if add_to_repo():
print "Copied data to pandora_motion_control"
else:
pass