def _render_actions_probs(self, surface, players, action_probs):
direction_to_rotation = {Direction.NORTH:0, Direction.WEST:90 , Direction.SOUTH:180, Direction.EAST:270}
direction_to_aligns = {
Direction.NORTH: {"horizontal_align": "center", "vertical_align":"bottom"},
Direction.WEST: {"horizontal_align": "right", "vertical_align":"center"},
Direction.SOUTH: {"horizontal_align": "center", "vertical_align":"top"},
Direction.EAST: {"horizontal_align": "left", "vertical_align":"center"}}
rescaled_arrow = pygame.transform.scale(self.ARROW_IMG, (self.tile_size, self.tile_size))
# divide width by math.sqrt(2) to always fit both interact icon and stay icon into single tile
rescaled_interact = pygame.transform.scale(self.INTERACT_IMG, (int(self.tile_size/math.sqrt(2)), self.tile_size))
rescaled_stay = pygame.transform.scale(self.STAY_IMG, (int(self.tile_size/math.sqrt(2)), self.tile_size))
for player, probs in zip(players, action_probs):
if probs is not None:
for action in Action.ALL_ACTIONS:
# use math sqrt to make probability proportional to area of the image
size = math.sqrt(probs[Action.ACTION_TO_INDEX[action]])
if action == "interact":
img = pygame.transform.rotozoom(rescaled_interact, 0, size)
self._render_on_tile_position(surface, img, player.position, horizontal_align="left", vertical_align="center")
elif action == Action.STAY:
img = pygame.transform.rotozoom(rescaled_stay, 0, size)
self._render_on_tile_position(surface, img, player.position, horizontal_align="right", vertical_align="center")
else:
position = Action.move_in_direction(player.position, action)
img = pygame.transform.rotozoom(rescaled_arrow, direction_to_rotation[action], size)
self._render_on_tile_position(surface, img, position, **direction_to_aligns[action])
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 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 get_near_locations(self, location):
"""Get neighbouring locations to the passed in location"""
near_locations = []
for d in Direction.ALL_DIRECTIONS:
new_location = Action.move_in_direction(location, d)
if self.is_in_bounds(new_location):
near_locations.append(new_location)
return near_locations
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)
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}