def actions(self, states, agent_indices):
action_probs_n = self.policy.multi_state_policy(states, agent_indices)
actions_and_infos_n = []
for action_probs in action_probs_n:
action = Action.sample(action_probs)
actions_and_infos_n.append((action, {"action_probs": action_probs}))
return actions_and_infos_n
def action(self, state):
action_probs = np.zeros(Action.NUM_ACTIONS)
legal_actions = list(Action.MOTION_ACTIONS)
if self.interact:
legal_actions.append(Action.INTERACT)
legal_actions_indices = np.array([Action.ACTION_TO_INDEX[motion_a] for motion_a in legal_actions])
action_probs[legal_actions_indices] = 1 / len(legal_actions_indices)
return Action.sample(action_probs), {"action_probs": action_probs}
def action(self, state):
action_probs = np.zeros(Action.NUM_ACTIONS)
legal_actions = list(Action.MOTION_ACTIONS)
if self.all_actions:
legal_actions = Action.ALL_ACTIONS
legal_actions_indices = np.array([Action.ACTION_TO_INDEX[motion_a] for motion_a in legal_actions])
action_probs[legal_actions_indices] = 1 / len(legal_actions_indices)
if self.custom_wait_prob is not None:
stay = Action.STAY
if np.random.random() < self.custom_wait_prob:
return stay, {"action_probs": Agent.a_probs_from_action(stay)}
else:
action_probs = Action.remove_indices_and_renormalize(action_probs, [Action.ACTION_TO_INDEX[stay]])
return Action.sample(action_probs), {"action_probs": action_probs}
def action(self, state):
action_probs = np.zeros(Action.NUM_ACTIONS)
for agent in self.agents:
action_probs += agent.action(state)[1]["action_probs"]
action_probs = action_probs / len(self.agents)
return Action.sample(action_probs), {"action_probs": action_probs}