def step(self, state, random=False):
if random:
action = self.agent.get_random_action()
else:
action = self.agent.act(state)
next_state = self.physics.step(GridActions(action))
reward = self.rewards.calculate_reward(state, GridActions(action), next_state)
self.trainer.add_experience(state, action, reward, next_state)
self.stats.add_experience((state, action, reward, copy.deepcopy(next_state)))
self.step_count += 1
return copy.deepcopy(next_state)
def test_episode(self, scenario=None):
state = copy.deepcopy(self.init_episode(scenario))
self.stats.on_episode_begin(self.episode_count)
while not state.terminal:
action = self.agent.get_exploitation_action_target(state)
next_state = self.physics.step(GridActions(action))
reward = self.rewards.calculate_reward(state, GridActions(action), next_state)
self.stats.add_experience((copy.deepcopy(state), action, reward, copy.deepcopy(next_state)))
state = copy.deepcopy(next_state)
self.stats.on_episode_end(self.episode_count)
self.stats.log_testing_data(step=self.step_count)
def test_scenario(self, scenario):
state = copy.deepcopy(self.init_episode(scenario))
while not state.all_terminal:
for state.active_agent in range(state.num_agents):
if state.terminal:
continue
action = self.agent.get_exploitation_action_target(state)
state = self.physics.step(GridActions(action))
def step(self, state: DHState, random=False):
for state.active_agent in range(state.num_agents):
if state.terminal:
continue
if random:
action = self.agent.get_random_action()
else:
action = self.agent.act(state)
if not self.first_action:
reward = self.rewards.calculate_reward(
self.last_states[state.active_agent],
GridActions(self.last_actions[state.active_agent]), state)
self.trainer.add_experience(
self.last_states[state.active_agent],
self.last_actions[state.active_agent], reward, state)
self.stats.add_experience(
(self.last_states[state.active_agent],
self.last_actions[state.active_agent], reward,
copy.deepcopy(state)))
self.last_states[state.active_agent] = copy.deepcopy(state)
self.last_actions[state.active_agent] = action
state = self.physics.step(GridActions(action))
if state.terminal:
reward = self.rewards.calculate_reward(
self.last_states[state.active_agent],
GridActions(self.last_actions[state.active_agent]), state)
self.trainer.add_experience(
self.last_states[state.active_agent],
self.last_actions[state.active_agent], reward, state)
self.stats.add_experience(
(self.last_states[state.active_agent],
self.last_actions[state.active_agent], reward,
copy.deepcopy(state)))
self.step_count += 1
self.first_action = False
return state
def test_episode(self):
state = copy.deepcopy(self.init_episode())
self.stats.on_episode_begin(self.episode_count)
first_action = True
while not state.all_terminal:
for state.active_agent in range(state.num_agents):
if state.terminal:
continue
action = self.agent.get_exploitation_action_target(state)
if not first_action:
reward = self.rewards.calculate_reward(
self.last_states[state.active_agent],
GridActions(self.last_actions[state.active_agent]),
state)
self.stats.add_experience(
(self.last_states[state.active_agent],
self.last_actions[state.active_agent], reward,
copy.deepcopy(state)))
self.last_states[state.active_agent] = copy.deepcopy(state)
self.last_actions[state.active_agent] = action
state = self.physics.step(GridActions(action))
if state.terminal:
reward = self.rewards.calculate_reward(
self.last_states[state.active_agent],
GridActions(self.last_actions[state.active_agent]),
state)
self.stats.add_experience(
(self.last_states[state.active_agent],
self.last_actions[state.active_agent], reward,
copy.deepcopy(state)))
first_action = False
self.stats.on_episode_end(self.episode_count)
self.stats.log_testing_data(step=self.step_count)
def test_scenario(self, scenario):
state = copy.deepcopy(self.init_episode(scenario))
while not state.terminal:
action = self.agent.get_exploitation_action_target(state)
state = self.physics.step(GridActions(action))