def make_mdp(mdp_class="grid", grid_dim=7):
'''
Returns:
(MDP)
'''
# Grid/Hallway stuff.
width, height = grid_dim, grid_dim
hall_goal_locs = [(i, width) for i in range(1, height+1)]
four_room_goal_locs = [(width, height), (width, 1), (1, height), (1, height - 2), (width - 2, height - 2), (width - 2, 1)]
# four_room_goal_loc = four_room_goal_locs[5]
# Taxi stuff.
agent = {"x":1, "y":1, "has_passenger":0}
passengers = [{"x":grid_dim / 2, "y":grid_dim / 2, "dest_x":grid_dim-2, "dest_y":2, "in_taxi":0}]
walls = []
mdp = {"hall":GridWorldMDP(width=width, height=height, init_loc=(1, 1), goal_locs=hall_goal_locs),
"pblocks_grid":make_grid_world_from_file("pblocks_grid.txt", randomize=True),
"grid":GridWorldMDP(width=width, height=height, init_loc=(1, 1), goal_locs=[(grid_dim, grid_dim)]),
"four_room":FourRoomMDP(width=width, height=height, goal_locs=[four_room_goal_loc]),
"chain":ChainMDP(num_states=grid_dim),
"random":RandomMDP(num_states=50, num_rand_trans=2),
"hanoi":HanoiMDP(num_pegs=grid_dim, num_discs=3),
"taxi":TaxiOOMDP(width=grid_dim, height=grid_dim, slip_prob=0.0, agent=agent, walls=walls, passengers=passengers)}[mdp_class]
return mdp
def make_mdp(mdp_class="grid", grid_dim=7):
'''
Returns:
(MDP)
'''
# Grid/Hallway stuff.
width, height = grid_dim, grid_dim
hall_goal_locs = [(i, width) for i in range(1, height + 1)]
four_room_goal_locs = [(width, height), (width, 1), (1, height),
(1, height - 2), (width - 2, height - 2),
(width - 2, 1)]
four_room_goal_loc = four_room_goal_locs[5]
# Taxi stuff.
agent = {"x": 1, "y": 1, "has_passenger": 0}
passengers = [{
"x": grid_dim / 2,
"y": grid_dim / 2,
"dest_x": grid_dim - 2,
"dest_y": 2,
"in_taxi": 0
}]
walls = []
mdp = {
"hall":
GridWorldMDP(width=width,
height=height,
init_loc=(1, 1),
goal_locs=hall_goal_locs),
"pblocks_grid":
make_grid_world_from_file("pblocks_grid.txt", randomize=True),
"grid":
GridWorldMDP(width=width,
height=height,
init_loc=(1, 1),
goal_locs=[(grid_dim, grid_dim)]),
"four_room":
FourRoomMDP(width=width, height=height,
goal_locs=[four_room_goal_loc]),
"chain":
ChainMDP(num_states=grid_dim),
"random":
RandomMDP(num_states=50, num_rand_trans=2),
"hanoi":
HanoiMDP(num_pegs=grid_dim, num_discs=3),
"taxi":
TaxiOOMDP(width=grid_dim,
height=grid_dim,
slip_prob=0.0,
agent=agent,
walls=walls,
passengers=passengers)
}[mdp_class]
return mdp
def run_all_experiments():
# TODOs
# 1. Add Taxi
# 2. Add ???
domains = ['9x9grid', 'fourroom', 'Imaze']
for d in domains:
fname = d + '.txt'
mdp = make_grid_world_from_file(fname)
run_MIMO(domain, mdp, [1, 2, 3])
run_MOMI(domain, mdp, [12, 11, 10, 9, 8, 7, 6, 5, 4])
def make_mdp(mdp_class="grid", state_size=7):
'''
Returns:
(MDP)
'''
# Grid/Hallway stuff.
width, height = state_size, state_size
hall_goal_locs = [(i, width) for i in range(1, height + 1)]
# Taxi stuff.
agent = {"x": 1, "y": 1, "has_passenger": 0}
passengers = [{
"x": state_size / 2,
"y": state_size / 2,
"dest_x": state_size - 2,
"dest_y": 2,
"in_taxi": 0
}]
walls = []
mdp = {
"hall":
GridWorldMDP(width=width,
height=height,
init_loc=(1, 1),
goal_locs=hall_goal_locs),
"pblocks_grid":
make_grid_world_from_file("pblocks_grid.txt", randomize=True),
"grid":
GridWorldMDP(width=width,
height=height,
init_loc=(1, 1),
goal_locs=[(state_size, state_size)]),
"four_room":
FourRoomMDP(width=width, height=height, goal_locs=[(width, height)]),
"chain":
ChainMDP(num_states=state_size),
"random":
RandomMDP(num_states=50, num_rand_trans=2),
"taxi":
TaxiOOMDP(width=state_size,
height=state_size,
slip_prob=0.0,
agent=agent,
walls=walls,
passengers=passengers)
}[mdp_class]
return mdp
def TestMatching():
domain = '5x5grid'
fname = '../tasks/' + domain + '.txt'
mdp = make_grid_world_from_file(fname)
G, intToS = GetAdjacencyMatrix(mdp)
c = GetCost(G)
matrix, F, LB = MinimumWeightMatching(G, c)
print('F\'=', F)
print('LB=', LB)
Gnx = nx.from_edgelist(F)
dic = dict()
for i, s in enumerate(intToS):
dic[i] = (s.x, s.y)
nx.draw_networkx_nodes(Gnx, pos=dic, node_size=300, node_color='g')
nx.draw_networkx_edges(Gnx, pos=dic)
plt.savefig('Matching.pdf')
def get_mdp_params(args):
state_dim = None
state_bound = None
num_actions = None
action_dim = None
action_bound = None
# TODO: it is very hard to have a script which contains all
# discrete/continuous state/actions.
# Should we separete out the tasks, or refactor?
if args.tasktype == 'pinball' or args.tasktype == 'p':
# TODO: Add parameter for Configuration files by --task argument
mdp = PinballMDP(cfg=args.task, render=args.render)
state_dim = 4
num_actions = len(mdp.get_actions())
# assert(args.ffunction != 'fourier')
elif args.tasktype == 'atari' or args.tasktype == 'atariram':
grayscale = False
downscale = True
# downscale = args.tasktype == 'atari'
mdp = GymMDP(env_name=args.task, grayscale=grayscale, downscale=downscale, render=args.render)
# mdp = GymMDP(env_name=args.task, grayscale=True, render=args.render)
mdp.env.seed(1234)
state_dims = mdp.env.observation_space.shape
# print('observation_space=', state_dims)
if args.tasktype == 'atari':
state_dim = 1
for d in state_dims:
state_dim *= d
# state_dim = 33600
# state_dim = 40000 # ?
if grayscale:
state_dim = int(state_dim / 3)
if downscale:
# state_dim = int(state_dim / 4)
state_dim = 105 * 80 * 3
else:
state_dim = 128
print('state_dim=', state_dim)
num_actions = mdp.env.action_space.n
# TODO: methods are fixed to dqn/ddpg/nn right now.
print('args.highmethod is overwritten by dqn')
print('args.lowmethod is overwritten by dqn')
args.highmethod = 'dqn'
args.lowmethod = 'dqn'
# args.ffunction = 'nn'
assert(args.highmethod == 'dqn')
assert(args.lowmethod == 'dqn')
# assert(args.ffunction == 'nn')
elif args.tasktype == 'mujoco':
mdp = GymMDP(env_name=args.task, render=args.render)
mdp.env.seed(1234)
state_dims = mdp.env.observation_space.shape
state_dim = 1
for d in state_dims:
state_dim *= d
print('state_dim=', state_dim)
action_dim = int(mdp.env.action_space.shape[0])
action_bound = mdp.action_bounds()
# print(action_dim)
# Fourier does not work for high dim space.
# TODO: methods are fixed to dqn/ddpg/nn right now.
print('args.highmethod is overwritten by dqn')
print('args.lowmethod is overwritten by ddpg')
args.highmethod = 'dqn'
args.lowmethod = 'ddpg'
# args.ffunction = 'nn'
assert(args.highmethod == 'dqn')
assert(args.lowmethod == 'ddpg')
# assert(args.ffunction == 'nn')
pass
elif args.tasktype == 'grid':
fname = '../tasks/' + args.task
mdp = make_grid_world_from_file(fname)
state_dim = 2
num_actions = 4
else:
assert(False)
pass
state_bound = mdp.bounds()
return mdp, state_dim, state_bound, num_actions, action_dim, action_bound
if __name__ == "__main__":
TestMatching()
exit(0)
# domain = '5x5grid'
# goals = [(1, 5), (1, 1), (5, 5), (3, 3), (5, 1)]
domain = '9x9grid'
goals = [(1, 1), (1, 9), (9, 1), (9, 9), (5, 5)]
# domain = 'fourroom'
# goals = [(1, 1), (1, 11), (11, 1), (11, 11), (5, 5), (8, 7), (5, 7)]
fname = '../../tasks/' + domain + '.txt'
mdp = make_grid_world_from_file(fname)
G, intToS = GetAdjacencyMatrix(mdp)
c = np.ones_like(G, dtype=int)
d = GetCost(G)
# print('d=', d)
# TODO
K = StatesToArray(intToS, goals)
# K = np.random.binomial(n=1, p=0.2, size=G.shape[0]) # np.ones(G.shape[0], dtype=int)
print('K=', K)
D = 15
tree, options = DiameterConstrainedSteinerTree(G, c, d, K, D, 0.1)
parser.add_argument(
'--nsepisodes',
type=int,
default=10,
help='number of episodes for incidence matrix sampling')
parser.add_argument('--nssteps',
type=int,
default=10,
help='number of steps for incidence matrix sampling')
args = parser.parse_args()
dom, task = args.task.split('_')
if dom == 'grid':
mdp = make_grid_world_from_file('../tasks/' + task + '.txt')
elif dom == 'taxi':
width = 4
height = 4
agent = {"x": 1, "y": 1, "has_passenger": 0}
passengers = [{"x": 3, "y": 2, "dest_x": 2, "dest_y": 3, "in_taxi": 0}]
mdp = TaxiOOMDP(width, height, agent, walls=[], passengers=passengers)
elif dom == 'gym':
mdp = GymMDP(env_name=task, render=False)
elif dom == 'hanoi':
mdp = HanoiMDP(num_pegs=3, num_discs=4)
elif dom == 'track':
mdp = make_race_track_from_file('../tasks/' + task + '.txt')
else:
print('Unknown task name: ', task)
assert (False)
def make_mdp_distr(mdp_class="grid", grid_dim=7, horizon=0):
'''
Args:
mdp_class (str): one of {"grid", "random"}
horizon (int)
Returns:
(MDPDistribution)
'''
mdp_dist_dict = {}
height, width = grid_dim, grid_dim
# Define goal locations.
# Corridor.
corr_width = 20
corr_goal_magnitude = random.randint(1, 5)
corr_goal_cols = [i for i in xrange(1, corr_goal_magnitude)] + [
j for j in xrange(corr_width - corr_goal_magnitude, corr_width + 1)
]
corr_goal_locs = list(itertools.product(corr_goal_cols, [1]))
# Grid World
grid_world_rows, grid_world_cols = [i for i in xrange(width - 4, width)], [
j for j in xrange(height - 4, height)
]
grid_goal_locs = list(itertools.product(grid_world_rows, grid_world_cols))
# Hallway.
hall_goal_locs = [(i, width) for i in range(1, height + 1)]
# Four room.
four_room_goal_locs = [(2, 2), (width, height), (width, 1), (1, height)]
# Taxi.
agent = {"x": 1, "y": 1, "has_passenger": 0}
walls = []
goal_loc_dict = {
"four_room": four_room_goal_locs,
"hall": hall_goal_locs,
"grid": grid_goal_locs,
"corridor": corr_goal_locs
}
# MDP Probability.
num_mdps = 10 if mdp_class not in goal_loc_dict.keys() else len(
goal_loc_dict[mdp_class])
mdp_prob = 1.0 / num_mdps
for i in range(num_mdps):
new_mdp = {"hall":GridWorldMDP(width=width, height=height, init_loc=(1, 1), goal_locs=[goal_loc_dict["hall"][i % len(goal_loc_dict["hall"])]]),
"corridor":GridWorldMDP(width=20, height=1, init_loc=(10, 1), goal_locs=[goal_loc_dict["corridor"][i % len(goal_loc_dict["corridor"])]], is_goal_terminal=True),
"grid":GridWorldMDP(width=width, height=height, init_loc=(1, 1), goal_locs=[goal_loc_dict["grid"][i % len(goal_loc_dict["grid"])]], is_goal_terminal=True),
"four_room":FourRoomMDP(width=width, height=height, goal_locs=[goal_loc_dict["four_room"][i % len(goal_loc_dict["four_room"])]]),
# THESE GOALS ARE SPECIFIED IMPLICITLY:
"pblocks_grid":make_grid_world_from_file("pblocks_grid.txt", randomize=True),
"chain":ChainMDP(num_states=10, reset_val=random.choice([0, 0.01, 0.05, 0.1, 0.2, 0.5])),
"random":RandomMDP(num_states=40, num_rand_trans=random.randint(1,10)),
"taxi":TaxiOOMDP(4, 4, slip_prob=0.0, agent=agent, walls=walls, \
passengers=[{"x":2, "y":2, "dest_x":random.randint(1,4), "dest_y":random.randint(1,4), "in_taxi":0}])}[mdp_class]
mdp_dist_dict[new_mdp] = mdp_prob
return MDPDistribution(mdp_dist_dict, horizon=horizon)
def make_mdp_distr(mdp_class="grid",
grid_dim=9,
horizon=0,
step_cost=0,
gamma=0.99):
'''
Args:
mdp_class (str): one of {"grid", "random"}
horizon (int)
step_cost (float)
gamma (float)
Returns:
(MDPDistribution)
'''
mdp_dist_dict = {}
height, width = grid_dim, grid_dim
# Define goal locations.
# Corridor.
corr_width = 20
corr_goal_magnitude = 1 #random.randint(1, 5)
corr_goal_cols = [i for i in range(1, corr_goal_magnitude + 1)] + [
j for j in range(corr_width - corr_goal_magnitude + 1, corr_width + 1)
]
corr_goal_locs = list(itertools.product(corr_goal_cols, [1]))
# Grid World
tl_grid_world_rows, tl_grid_world_cols = [
i for i in range(width - 4, width)
], [j for j in range(height - 4, height)]
tl_grid_goal_locs = list(
itertools.product(tl_grid_world_rows, tl_grid_world_cols))
tr_grid_world_rows, tr_grid_world_cols = [i for i in range(1, 4)], [
j for j in range(height - 4, height)
]
tr_grid_goal_locs = list(
itertools.product(tr_grid_world_rows, tr_grid_world_cols))
grid_goal_locs = tl_grid_goal_locs + tr_grid_goal_locs
# Hallway.
hall_goal_locs = [(i, height) for i in range(1, 30)]
# Four room.
four_room_goal_locs = [(width, height), (width, 1), (1, height),
(1, height - 2),
(width - 2, height - 2)] #, (width - 2, 1)]
# Taxi.
agent = {"x": 1, "y": 1, "has_passenger": 0}
walls = []
goal_loc_dict = {
"four_room": four_room_goal_locs,
"hall": hall_goal_locs,
"grid": grid_goal_locs,
"corridor": corr_goal_locs,
}
# MDP Probability.
num_mdps = 10 if mdp_class not in goal_loc_dict.keys() else len(
goal_loc_dict[mdp_class])
if mdp_class == "octo":
num_mdps = 12
mdp_prob = 1.0 / num_mdps
for i in range(num_mdps):
new_mdp = {"hrooms":make_grid_world_from_file("hierarch_rooms.txt", num_goals=7, randomize=False),
"octo":make_grid_world_from_file("octogrid.txt", num_goals=12, randomize=False, goal_num=i),
"hall":GridWorldMDP(width=30, height=height, rand_init=False, goal_locs=goal_loc_dict["hall"], name="hallway", is_goal_terminal=True),
"corridor":GridWorldMDP(width=20, height=1, init_loc=(10, 1), goal_locs=[goal_loc_dict["corridor"][i % len(goal_loc_dict["corridor"])]], is_goal_terminal=True, name="corridor"),
"grid":GridWorldMDP(width=width, height=height, rand_init=True, goal_locs=[goal_loc_dict["grid"][i % len(goal_loc_dict["grid"])]], is_goal_terminal=True),
"four_room":FourRoomMDP(width=width, height=height, goal_locs=[goal_loc_dict["four_room"][i % len(goal_loc_dict["four_room"])]], is_goal_terminal=True),
# THESE GOALS ARE SPECIFIED IMPLICITLY:
"pblocks_grid":make_grid_world_from_file("pblocks_grid.txt", randomize=True, slip_prob=0.1),
"chain":ChainMDP(num_states=10, reset_val=random.choice([0, 0.01, 0.05, 0.1, 0.2, 0.5])),
"random":RandomMDP(num_states=40, num_rand_trans=random.randint(1,10)),
"taxi":TaxiOOMDP(3, 4, slip_prob=0.0, agent=agent, walls=walls, \
passengers=[{"x":2, "y":1, "dest_x":random.choice([2,3]), "dest_y":random.choice([2,3]), "in_taxi":0},
{"x":1, "y":2, "dest_x":random.choice([1,2]), "dest_y":random.choice([1,4]), "in_taxi":0}])}[mdp_class]
new_mdp.set_step_cost(step_cost)
new_mdp.set_gamma(gamma)
mdp_dist_dict[new_mdp] = mdp_prob
return MDPDistribution(mdp_dist_dict, horizon=horizon)
def make_mdp_distr(mdp_class="grid", grid_dim=9, horizon=0, step_cost=0, gamma=0.99):
'''
Args:
mdp_class (str): one of {"grid", "random"}
horizon (int)
step_cost (float)
gamma (float)
Returns:
(MDPDistribution)
'''
mdp_dist_dict = {}
height, width = grid_dim, grid_dim
# Define goal locations.
# Corridor.
corr_width = 20
corr_goal_magnitude = 1 #random.randint(1, 5)
corr_goal_cols = [i for i in xrange(1, corr_goal_magnitude + 1)] + [j for j in xrange(corr_width-corr_goal_magnitude + 1, corr_width + 1)]
corr_goal_locs = list(itertools.product(corr_goal_cols, [1]))
# Grid World
tl_grid_world_rows, tl_grid_world_cols = [i for i in xrange(width - 4, width)], [j for j in xrange(height - 4, height)]
tl_grid_goal_locs = list(itertools.product(tl_grid_world_rows, tl_grid_world_cols))
tr_grid_world_rows, tr_grid_world_cols = [i for i in xrange(1, 4)], [j for j in xrange(height - 4, height)]
tr_grid_goal_locs = list(itertools.product(tr_grid_world_rows, tr_grid_world_cols))
grid_goal_locs = tl_grid_goal_locs + tr_grid_goal_locs
# Hallway.
upworld_goal_locs = [(i, height) for i in xrange(1, 30)]
# Four room.
four_room_goal_locs = [(width, height), (width, 1), (1, height), (1, height - 2), (width - 2, height - 2), (width - 2, 1)]
print four_room_goal_locs
tight_four_room_goal_locs = [(width, height), (width, height-1), (width-1, height), (width, height - 2), (width - 2, height), (width-1, height-1)]
# Taxi.
agent = {"x":1, "y":1, "has_passenger":0}
walls = []
goal_loc_dict = {"four_room":four_room_goal_locs,
"color":four_room_goal_locs,
"upworld":upworld_goal_locs,
"grid":grid_goal_locs,
"corridor":corr_goal_locs,
"tight_four_room":tight_four_room_goal_locs,
}
# MDP Probability.
num_mdps = 10 if mdp_class not in goal_loc_dict.keys() else len(goal_loc_dict[mdp_class])
if mdp_class == "octo":
num_mdps = 12
mdp_prob = 1.0 / num_mdps
for i in xrange(num_mdps):
new_mdp = {"hrooms":make_grid_world_from_file("hierarch_rooms.txt", num_goals=7, randomize=False),
"octo":make_grid_world_from_file("octogrid.txt", num_goals=12, randomize=False, goal_num=i),
"upworld":GridWorldMDP(width=30, height=height, rand_init=False, goal_locs=goal_loc_dict["upworld"], name="upworld", is_goal_terminal=True),
"corridor":GridWorldMDP(width=20, height=1, init_loc=(10, 1), goal_locs=[goal_loc_dict["corridor"][i % len(goal_loc_dict["corridor"])]], is_goal_terminal=True, name="corridor"),
"grid":GridWorldMDP(width=width, height=height, rand_init=True, goal_locs=[goal_loc_dict["grid"][i % len(goal_loc_dict["grid"])]], is_goal_terminal=True),
"four_room":FourRoomMDP(width=width, height=height, goal_locs=[goal_loc_dict["four_room"][i % len(goal_loc_dict["four_room"])]], is_goal_terminal=True),
"color":ColorMDP(width=width, height=height, num_colors=4, goal_locs=[goal_loc_dict["four_room"][i % len(goal_loc_dict["four_room"])]], is_goal_terminal=True),
"tight_four_room":FourRoomMDP(width=width, height=height, goal_locs=[goal_loc_dict["tight_four_room"][i % len(goal_loc_dict["tight_four_room"])]], is_goal_terminal=True, name="tight_four_room")}[mdp_class]
new_mdp.set_step_cost(step_cost)
new_mdp.set_gamma(gamma)
mdp_dist_dict[new_mdp] = mdp_prob
return MDPDistribution(mdp_dist_dict, horizon=horizon)
parser.add_argument(
'--nsepisodes',
type=int,
default=10,
help='number of episodes for incidence matrix sampling')
parser.add_argument('--nssteps',
type=int,
default=10,
help='number of steps for incidence matrix sampling')
args = parser.parse_args()
dom, task = args.task.split('_')
if dom == 'grid':
mdp = make_grid_world_from_file('options/tasks/' + task + '.txt')
elif dom == 'taxi':
width = 4
height = 4
agent = {"x": 1, "y": 1, "has_passenger": 0}
passengers = [{"x": 3, "y": 2, "dest_x": 2, "dest_y": 3, "in_taxi": 0}]
mdp = TaxiOOMDP(width, height, agent, walls=[], passengers=passengers)
elif dom == 'gym':
mdp = GymMDP(env_name=task, render=False)
elif dom == 'hanoi':
mdp = HanoiMDP(num_pegs=3, num_discs=4)
elif dom == 'track':
mdp = make_race_track_from_file('../tasks/' + task + '.txt')
else:
print('Unknown task name: ', task)
assert (False)
def make_mdp_distr(mdp_class, is_goal_terminal, mdp_size=11, horizon=0, gamma=0.99):
'''
Args:
mdp_class (str): one of {"grid", "random"}
horizon (int)
step_cost (float)
gamma (float)
Returns:
(MDPDistribution)
'''
mdp_dist_dict = {}
height, width, = mdp_size, mdp_size
# Corridor.
corr_width = 20
corr_goal_magnitude = 1 #random.randint(1, 5)
corr_goal_cols = [i for i in xrange(1, corr_goal_magnitude + 1)] + [j for j in xrange(corr_width-corr_goal_magnitude + 1, corr_width + 1)]
corr_goal_locs = list(itertools.product(corr_goal_cols, [1]))
# Grid World
tl_grid_world_rows, tl_grid_world_cols = [i for i in xrange(width - 4, width)], [j for j in xrange(height - 4, height)]
tl_grid_goal_locs = list(itertools.product(tl_grid_world_rows, tl_grid_world_cols))
tr_grid_world_rows, tr_grid_world_cols = [i for i in xrange(1, 4)], [j for j in xrange(height - 4, height)]
tr_grid_goal_locs = list(itertools.product(tr_grid_world_rows, tr_grid_world_cols))
grid_goal_locs = tl_grid_goal_locs + tr_grid_goal_locs
# Four room.
four_room_goal_locs = [(width, height), (width, 1), (1, height), (1, height - 2), (width - 2, height - 2)]#, (width - 2, 1)]
# SPREAD vs. TIGHT
spread_goal_locs = [(width, height), (width, 1), (1, height), (1, height - 2), (width - 2, height - 2), (width - 2, 1), (2,2)]
tight_goal_locs = [(width, height), (width-1, height), (width, height-1), (width, height - 2), (width - 2, height), (width - 1, height-1), (width-2,height-2)]
changing_entities = {"four_room":four_room_goal_locs,
"grid":grid_goal_locs,
"corridor":corr_goal_locs,
"spread":spread_goal_locs,
"tight":tight_goal_locs,
"chain":[0.0, 0.01, 0.1, 0.5, 1.0],
"combo_lock":[[3,1,2],[3,2,1],[2,3,1],[3,3,1]],
"walls":make_wall_permutations(mdp_size),
"lava":make_lava_permutations(mdp_size)
}
# MDP Probability.
num_mdps = 10 if mdp_class not in changing_entities.keys() else len(changing_entities[mdp_class])
if mdp_class == "octo":
num_mdps = 12
mdp_prob = 1.0 / num_mdps
for i in xrange(num_mdps):
new_mdp = {"chain":ChainMDP(reset_val=changing_entities["chain"][i%len(changing_entities["chain"])]),
"lava":GridWorldMDP(width=width, height=height, rand_init=False, lava_locs=changing_entities["lava"][i%len(changing_entities["lava"])], goal_locs=[(mdp_size-3, mdp_size-3)], is_goal_terminal=is_goal_terminal, name="lava_world", slip_prob=0.1),
"four_room":FourRoomMDP(width=width, height=height, goal_locs=[changing_entities["four_room"][i % len(changing_entities["four_room"])]], is_goal_terminal=is_goal_terminal),
"octo":make_grid_world_from_file("octogrid.txt", num_goals=12, randomize=False, goal_num=i),
"corridor":GridWorldMDP(width=20, height=1, init_loc=(10, 1), goal_locs=[changing_entities["corridor"][i % len(changing_entities["corridor"])]], is_goal_terminal=is_goal_terminal, name="corridor"),
"walls":ThinWallGridMDP(width=width, height=height, walls=changing_entities["walls"][i%len(changing_entities["walls"])]),
"combo_lock":ComboLockMDP(combo=changing_entities["combo_lock"][i%len(changing_entities["combo_lock"])]),
"spread":GridWorldMDP(width=width, height=height, rand_init=False, goal_locs=[changing_entities["spread"][i % len(changing_entities["spread"])]], is_goal_terminal=is_goal_terminal, name="spread_grid"),
"tight":GridWorldMDP(width=10, height=10, rand_init=False, goal_locs=[changing_entities["tight"][i % len(changing_entities["tight"])]], is_goal_terminal=is_goal_terminal, name="tight_grid"),
}[mdp_class]
new_mdp.set_gamma(gamma)
mdp_dist_dict[new_mdp] = mdp_prob
return MDPDistribution(mdp_dist_dict, horizon=horizon)
'--nsepisodes',
type=int,
default=10,
help='number of episodes for incidence matrix sampling')
parser.add_argument('--nssteps',
type=int,
default=10,
help='number of steps for incidence matrix sampling')
args = parser.parse_args()
dom, task = args.task.split('_')
if dom == 'grid':
mdp = make_grid_world_from_file(
os.path.dirname(os.path.realpath(__file__)) + '/../tasks/' + task +
'.txt')
elif dom == 'taxi':
width = 4
height = 4
agent = {"x": 1, "y": 1, "has_passenger": 0}
passengers = [{"x": 3, "y": 2, "dest_x": 2, "dest_y": 3, "in_taxi": 0}]
mdp = TaxiOOMDP(width, height, agent, walls=[], passengers=passengers)
elif dom == 'gym':
mdp = GymMDP(env_name=task, render=False)
elif dom == 'hanoi':
mdp = HanoiMDP(num_pegs=3, num_discs=4)
elif dom == 'track':
mdp = make_race_track_from_file(
os.path.dirname(os.path.realpath(__file__)) + '/../tasks/' + task +
'.txt')
