def get_cliff(gamma):
ascii_room = """
#######
# !!! #
# #
# #
# #
#######"""[1:].split('\n')
ascii_room = [row.strip() for row in ascii_room]
# a = list(ascii_room[goal[0]])
# a[goal[1]] = "g"
# ascii_room[goal[0]] = "".join(a)
char_matrix = get_char_matrix(ascii_room)
grid_size = len(char_matrix[0])
reward_spec = {(1, grid_size - 2): 1}
builder = emdp.gridworld.builder_tools.TransitionMatrixBuilder(
grid_size=grid_size, has_terminal_state=False)
R = create_reward_matrix(builder.P.shape[0],
builder.grid_size,
reward_spec,
action_space=builder.P.shape[1])
walls, empty_, bomb = ascii_to_walls(char_matrix) # hacks
empty = []
for e in empty_:
(e, ), = flatten_state(e, grid_size, grid_size * grid_size).nonzero()
# R[e, :] = -1
empty.append(e)
builder.add_grid(p_success=1, terminal_states=[(1, grid_size - 2)])
for (r, c) in walls:
builder.add_wall_at((r, c))
P = builder.P
for (r, c) in bomb:
idx = grid_size * r + c
R[idx, 0] = -100 # left
R[idx, 1] = -100 # rigth
R[idx, 2] = -100 # up
R[idx, 3] = -100 # down
# R[idx, 1] = -100
P[idx, :, :] = 0
P[idx, :, 8] = 1
p0 = np.zeros(R.shape[0])
p0[flatten_state((1, 1), grid_size, grid_size * grid_size).nonzero()] = 1.
# p0[empty] = 1 / len(empty)
gw = GridWorldMDP(P,
R,
gamma,
p0,
terminal_states=[(1, grid_size - 2)],
size=builder.grid_size)
return gw
def four_states_gw(goal):
builder = emdp.gridworld.builder_tools.TransitionMatrixBuilder(
grid_size=3, has_terminal_state=False)
builder.add_grid([], p_success=1)
builder.add_wall_at((0, 0))
builder.add_wall_at((0, 1))
builder.add_wall_at((0, 2))
builder.add_wall_at((1, 0))
builder.add_wall_at((1, 2))
reward_spec = {goal: +1}
R = create_reward_matrix(builder.P.shape[0],
builder.grid_size,
reward_spec,
action_space=builder.P.shape[1])
# target_state = pos_to_state(builder, (2, 2))
p0 = flatten_state((1, 1), builder.grid_size, R.shape[0])
# p0[[2, 4, 5, 8]] = .25
gw = GridWorldMDP(builder.P,
R,
0.9,
p0,
terminal_states=(),
size=builder.grid_size)
return gw
def four_rooms_gw(gamma):
ascii_room = """
##########
# ! ! #
# ! # !#
# ! #
### ! #
# !! #
# !! !#
# ! ! ! #
# ! #
##########"""[1:].split('\n')
ascii_room = [row.strip() for row in ascii_room]
char_matrix = get_char_matrix(ascii_room)
grid_size = len(char_matrix[0])
reward_spec = {(grid_size - 2, grid_size - 2): +1}
builder = emdp.gridworld.builder_tools.TransitionMatrixBuilder(
grid_size=grid_size, has_terminal_state=False)
walls, empty_, bomb = ascii_to_walls(char_matrix) # hacks
empty = []
for e in empty_:
(e, ), = flatten_state(e, grid_size, grid_size * grid_size).nonzero()
empty.append(e)
builder.add_grid(p_success=1,
terminal_states=[(grid_size - 2, grid_size - 2)])
for (r, c) in walls:
builder.add_wall_at((r, c))
for r, c in bomb:
builder.add_wall_at((r, c))
R = create_reward_matrix(builder.P.shape[0],
builder.grid_size,
reward_spec,
action_space=builder.P.shape[1])
#for b in bomb:
# idx = grid_size * b[0] + b[1]
# R[idx, :] = -10
p0 = np.zeros(R.shape[0])
p0[grid_size + 1] = 1
gw = GridWorldMDP(builder.P,
R,
gamma,
p0,
terminal_states=[(grid_size - 2, grid_size - 2)],
size=builder.grid_size)
return gw