def get_greedy_action(self, obs):
"""
Retrieves the best next action for the current observation according to the eval Deep Q-Network prediction.
:param obs: Current observation (state representation)
:return: action: AgentAction which should be chosen next by the agent according to the eval Deep Q-Network
"""
action_values = self.predict(np.array([obs]))[0]
greedy_action_index = np.argmax(action_values)
return index_to_agent_action(greedy_action_index)
def get_greedy_action(self, obs):
"""
Retrieves the best next action for the current observation according to the eval Deep Q-Network prediction.
:param obs: Current observation (state representation)
:return: action: AgentAction which should be chosen next by the agent according to the eval Deep Q-Network
"""
borda_count = self.compute_borda_count([obs])[0]
action_index = np.argmax(borda_count)
return index_to_agent_action(action_index)
def get_warm_up_action(self):
"""
Retrieves the action which should be played by the agent in the warm up phase (based on current turn number).
:return action : AgentAction which should be chosen next by the agent
"""
# Agents' request sequence is defined in agent_rule_requests
if self.turn < len(agent_rule_requests):
raw_action = agent_rule_requests[self.turn]
else:
raw_action = agent_rule_requests[-1]
feasible_action_index = raw_agent_action_to_index(raw_action)
action = index_to_agent_action(feasible_action_index)
return action
def choose_action(self, obs, warm_up=False):
"""
Determines which action is chosen for the next turn given the current observation.
Action choice can differ from greedy action policy when:
- A random action is chosen due to epsilon greedy exploration policy
- A predefined action sequence is used in the warm up phase
:param obs : Current observation (state representation)
:param warm_up : Flag whether the warm up phase is currently used
:return action : AgentAction which should be chosen next by the agent
"""
if warm_up:
action = self.get_warm_up_action()
elif random.random() < self.epsilon:
action_index = random.randrange(self.n_actions)
action = index_to_agent_action(action_index)
else:
action = self.get_greedy_action(obs)
action.round_num = self.turn
self.turn += 1
return action
def get_greedy_action(self, obs):
action_index = random.randrange(self.n_actions)
return index_to_agent_action(action_index)