def set_room(self, room):
assert 0 <= room < len(self.rooms)
self.room = room
roomarr = self.rooms[room]
self.states = roomarr.copy()
self.goal = filter_states(self.states, self.field.DOOR)[0]
self.initPos = filter_states(self.states, self.field.INIT)
self.lock_pos = filter_states(self.states, self.field.LOCK)[0]
key_pos = filter_states(self.states, self.field.KEY)
self.key_pos = key_pos[0] if key_pos else None
self.agent_holding_key = False
def perform_action(self, state, action):
assert action.shape == (1,)
action = action[0] # in this case action i s
pos2 = (state.x, state.y)
a = self.actions
reward = self.step_reward
if action in [a.N, a.S, a.E, a.W]:
delta = self.movemap[action]
pos2 = tuple(map(sum,zip(pos2, delta)))
if self.forbidden_pos(pos2):
pos2 = (state.x,state.y)
if pos2 == self.goal:
state = State(*[0]*17)._replace(x=pos2[0], y=pos2[1])
if not self.episode_finished(state):
self.set_room(self.room + 1)
reward = self.task_reward
state2 = self.new_state()
pos2 = (state2.x, state2.y)
else:
reward = self.final_reward
elif action == a.G:
if self.states[pos2] == self.field.KEY:
self.agent_holding_key = True
self.states[pos2] = self.field.EMPTY
reward = self.task_reward
elif action == a.P:
if (self.states[pos2] == self.field.LOCK and
not filter_states(self.states, self.field.KEY)):
self.agent_holding_key = False
self.states[self.goal] = self.field.EMPTY
#self.states[self.states == self.field.DOOR] = self.field.EMPTY
reward = self.task_reward
return (self.calculate_state(pos2), reward)
def choose_action_parameterized(self, env, state, field, action):
if self.room != state.r:
self.room = state.r
self.plan = None
pos = (state.x, state.y)
if self.plan is None or pos not in self.plan:
key_pos = filter_states(env.states, field)
# Not positive if this is right move here.
if not key_pos:
#print 'RANDOM action for plan', action
return np.array([choice(env.actions)])
priority_func = lambda s: manhattan_dist(s, key_pos[0])
expand_partial = lambda s: expand_state(env.states, s)
self.plan = search.best_first_graph_search(pos, key_pos[0], priority_func, expand_partial)
for i, pathpos in enumerate(self.plan):
if i == len(self.plan)-1:
#print 'action', action, 'for plan', field
return np.array([action])
elif pos == pathpos:
fx,fy = self.plan[i+1]
dx,dy = (fx-state.x, fy-state.y)
#print 'move', dx,dy, 'for plan', field
return np.array([env.movemap[dx,dy]])
