gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2026-04-01 01:23:33 +02:00
Code Issues Projects Releases Wiki Activity

multiple bug fixes in dealing with measurements + CartPole_DFP preset (#92)

Browse Source
This commit is contained in:
Itai Caspi
2018-04-23 10:44:46 +03:00
committed by GitHub
parent 5d5562bf62
commit 52eb159f69
5 changed files with 31 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
agents/agent.py
View File

@@ -365,7 +365,10 @@ class Agent(object):
'observation': observation
}
if self.tp.agent.use_measurements:
self.curr_state['measurements'] = self.env.measurements
if 'measurements' in self.curr_state.keys():
self.curr_state['measurements'] = self.env.state['measurements']
else:
self.curr_state['measurements'] = np.zeros(0)
if self.tp.agent.use_accumulated_reward_as_measurement:
self.curr_state['measurements'] = np.append(self.curr_state['measurements'], 0)
@@ -398,7 +401,7 @@ class Agent(object):
shaped_reward += action_info['action_intrinsic_reward']
# TODO: should total_reward_in_current_episode include shaped_reward?
self.total_reward_in_current_episode += result['reward']
next_state = result['state']
next_state = copy.copy(result['state'])
next_state['observation'] = self.preprocess_observation(next_state['observation'])
# plot action values online
@@ -411,8 +414,11 @@ class Agent(object):
observation = LazyStack(self.curr_stack, -1)
next_state['observation'] = observation
if self.tp.agent.use_measurements and 'measurements' in result.keys():
next_state['measurements'] = result['state']['measurements']
if self.tp.agent.use_measurements:
if 'measurements' in result['state'].keys():
next_state['measurements'] = result['state']['measurements']
else:
next_state['measurements'] = np.zeros(0)
if self.tp.agent.use_accumulated_reward_as_measurement:
next_state['measurements'] = np.append(next_state['measurements'], self.total_reward_in_current_episode)

9
agents/dfp_agent.py
View File

@@ -31,7 +31,7 @@ class DFPAgent(Agent):
# create the inputs for the network
input = current_states
input.append(np.repeat(np.expand_dims(self.current_goal, 0), self.tp.batch_size, 0))
input['goal'] = np.repeat(np.expand_dims(self.current_goal, 0), self.tp.batch_size, 0)
# get the current outputs of the network
targets = self.main_network.online_network.predict(input)
@@ -40,7 +40,7 @@ class DFPAgent(Agent):
for i in range(self.tp.batch_size):
targets[i, actions[i]] = batch[i].info['future_measurements'].flatten()
result = self.main_network.train_and_sync_networks(current_states, targets)
result = self.main_network.train_and_sync_networks(input, targets)
total_loss = result[0]
return total_loss
@@ -52,7 +52,10 @@ class DFPAgent(Agent):
goal = np.expand_dims(self.current_goal, 0)
# predict the future measurements
measurements_future_prediction = self.main_network.online_network.predict([observation, measurements, goal])[0]
measurements_future_prediction = self.main_network.online_network.predict({
"observation": observation,
"measurements": measurements,
"goal": goal})[0]
action_values = np.zeros((self.action_space_size,))
num_steps_used_for_objective = len(self.tp.agent.future_measurements_weights)