def make_mdp(mdp_class="grid", grid_dim=7):
'''
Returns:
(MDP)
'''
# Grid/Hallway stuff.
width, height = grid_dim, grid_dim
upworld_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 - 1, height - 1), (width - 2, 1)]
four_room_goal_loc = four_room_goal_locs[0]
# 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 = []
# Trench stuff
tr_agent = {"x": 1, "y": 1, "dx": 1, "dy": 0, "dest_x": grid_dim, "dest_y": grid_dim, "has_block": 0}
blocks = [{"x": grid_dim, "y": 1}]
lavas = [{"x": x, "y": y} for x, y in map(lambda z: (z + 1, (grid_dim + 1) / 2), range(grid_dim))]
# Do grids separately to avoid making error-prone domains.
if mdp_class == "four_room":
mdp = FourRoomMDP(width=width, height=height, goal_locs=[four_room_goal_loc])
else:
mdp = {"upworld":GridWorldMDP(width=width, height=height, init_loc=(1, 1), goal_locs=upworld_goal_locs),
"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, agent=agent, walls=walls, passengers=passengers),
"trench":TrenchOOMDP(width=grid_dim, height=3, agent=tr_agent, blocks=blocks, lavas=lavas)}[mdp_class]
return mdp
def choose_mdp(mdp_name, env_name="Asteroids-v0"):
'''
Args:
mdp_name (str): one of {gym, grid, chain, taxi, ...}
gym_env_name (str): gym environment name, like 'CartPole-v0'
Returns:
(MDP)
'''
# Other imports
from simple_rl.tasks import ChainMDP, GridWorldMDP, FourRoomMDP, TaxiOOMDP, RandomMDP, PrisonersDilemmaMDP, RockPaperScissorsMDP, GridGameMDP
# Taxi MDP.
agent = {"x":1, "y":1, "has_passenger":0}
passengers = [{"x":4, "y":3, "dest_x":2, "dest_y":2, "in_taxi":0}]
walls = []
if mdp_name == "gym":
# OpenAI Gym MDP.
try:
from simple_rl.tasks.gym.GymMDPClass import GymMDP
except:
raise ValueError("(simple_rl) Error: OpenAI gym not installed.")
return GymMDP(env_name, render=True)
else:
return {"grid":GridWorldMDP(5, 5, (1, 1), goal_locs=[(5, 3), (4,1)]),
"four_room":FourRoomMDP(),
"chain":ChainMDP(5),
"taxi":TaxiOOMDP(10, 10, slip_prob=0.0, agent=agent, walls=walls, passengers=passengers),
"random":RandomMDP(num_states=40, num_rand_trans=20),
"prison":PrisonersDilemmaMDP(),
"rps":RockPaperScissorsMDP(),
"grid_game":GridGameMDP(),
"multi":{0.5:RandomMDP(num_states=40, num_rand_trans=20), 0.5:RandomMDP(num_states=40, num_rand_trans=5)}}[mdp_name]
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 main():
# Make MDP.
grid_dim = 11
mdp = FourRoomMDP(width=grid_dim, height=grid_dim, init_loc=(1, 1), slip_prob=0.05, goal_locs=[(grid_dim, grid_dim)], gamma=0.99)
# Experiment Type.
exp_type = "learn_w_abstr"
# For comparing policies and visualizing.
beta = 1
is_deterministic_ib = True
is_agent_in_control = True
# For main plotting experiment.
beta_range = list(chart_utils.drange(0.0, 4.0, 1.0))
instances = 1
# Get demo policy.
vi = ValueIteration(mdp)
_, val = vi.run_vi()
# Epsilon greedy policy
demo_policy = get_lambda_policy(make_det_policy_eps_greedy(vi.policy, vi.get_states(), mdp.get_actions(), epsilon=0.1))
if exp_type == "plot_info_sa_val_and_num_states":
# Makes the main two plots.
make_info_sa_val_and_size_plots(mdp, demo_policy, beta_range, instances=instances, is_agent_in_control=is_agent_in_control)
elif exp_type == "compare_policies":
# Makes a plot comparing value of pi-phi combo from info_sa with \pi_d.
info_sa_compare_policies(mdp, demo_policy, beta=beta, is_deterministic_ib=is_deterministic_ib, is_agent_in_control=is_agent_in_control)
elif exp_type == "visualize_info_sa_abstr":
# Visualize the state abstraction found by info_sa.
info_sa_visualize_abstr(mdp, demo_policy, beta=beta, is_deterministic_ib=is_deterministic_ib, is_agent_in_control=is_agent_in_control)
elif exp_type == "learn_w_abstr":
# Run learning experiments for different settings of \beta.
learn_w_abstr(mdp, demo_policy, is_deterministic_ib=is_deterministic_ib)
elif exp_type == "planning":
info_sa_planning_experiment()
def main(open_plot=True):
# Setup MDP, Agents.
# mdp = GridWorldMDP(width=4, height=3, init_loc=(1,1), goal_locs=[(4,3)], gamma=0.95, walls=[(2,2)])
mdp = FourRoomMDP(width=11,
height=11,
init_loc=(1, 1),
goal_locs=[(9, 3)],
is_goal_terminal=True,
slip_prob=0.2)
# mdp = ComboLockMDP(combo=[3,1,2], num_actions=3, num_states=3)
dq_agent = DoubleQAgent(actions=mdp.get_actions())
ql_agent = QLearningAgent(actions=mdp.get_actions())
rand_agent = RandomAgent(actions=mdp.get_actions())
# Run experiment and make plot.
run_agents_on_mdp([ql_agent, rand_agent],
mdp,
instances=5,
episodes=1,
steps=10000,
open_plot=open_plot)
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 main():
# Setup MDP, Agents.
mdp = FourRoomMDP(9, 9, goal_locs=[(9, 9)], gamma=0.95)
ql_agent = QLearnerAgent(mdp.get_actions())
viz = parse_args()
if viz == "value":
# Run experiment and make plot.
mdp.visualize_value()
elif viz == "policy":
# Viz policy
vi = ValueIteration(mdp)
vi.run_vi()
policy = vi.policy
mdp.visualize_policy(policy)
elif viz == "agent":
# Solve problem and show agent interaction.
print "\n", str(ql_agent), "interacting with", str(mdp)
run_single_agent_on_mdp(ql_agent, mdp, episodes=500, steps=200)
mdp.visualize_agent(ql_agent)
def make_mdp_distr(mdp_class="grid", num_mdps=15, gamma=0.99):
'''
Args:
mdp_class (str): one of {"grid", "random"}
num_mdps (int)
Returns:
(MDPDistribution)
'''
mdp_dist_dict = {}
mdp_prob = 1.0 / num_mdps
height, width = 10, 10
# Make @num_mdps MDPs.
for i in xrange(num_mdps):
next_goals = rnd.sample([(1, 7), (7, 1), (7, 7), (6, 6), (6, 1),
(1, 6)], 2)
new_mdp = {
"grid":
GridWorldMDP(width=width,
height=height,
init_loc=(1, 1),
goal_locs=rnd.sample(
zip(range(1, width + 1), [height] * width), 1),
is_goal_terminal=True,
gamma=gamma),
"four_room":
FourRoomMDP(width=8, height=8, goal_locs=next_goals, gamma=gamma),
"chain":
ChainMDP(num_states=10,
reset_val=rnd.choice([0, 0.01, 0.05, 0.1]),
gamma=gamma),
"random":
RandomMDP(num_states=40,
num_rand_trans=rnd.randint(1, 10),
gamma=gamma)
}[mdp_class]
mdp_dist_dict[new_mdp] = mdp_prob
return MDPDistribution(mdp_dist_dict)
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)
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 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
# 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 range(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, step_cost=-0.001, lava_cost=0.0, 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"),
"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=width, height=height, 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)
def branching_factor_experiment(min_options=0,
max_options=20,
increment=2,
instances=5,
epsilon=0.05):
'''
Args:
min_options (int)
max_options (int)
increment (int)
Summary:
Runs an experiment contrasting learning performance for different # options.
'''
# Define MDP.
grid_size = 7
mdp = FourRoomMDP(width=grid_size,
height=grid_size,
goal_locs=[(grid_size, grid_size)])
# Make State Abstraction.
states, _ = ah.compute_reachable_state_space(mdp, sample_rate=50)
state_abstr = core.compute_phi_given_m(mdp,
four_rooms_predicate_9x9,
level=1,
states=states)
x_axis = range(min_options, max_options + 1, increment)
y_axis = defaultdict(list) #[] #[0] * len(x_axis)
conf_intervals = defaultdict(list)
num_options_performance = defaultdict(lambda: defaultdict(list))
# Choose dependent variable (either #steps per episode or #episodes).
d_var_range = [(20, 5), (40, 250), (400, 2500)]
for steps, episodes in d_var_range:
print "steps, episodes", steps, episodes
# Evaluate.
for i, instance in enumerate(range(instances)):
print "\tInstance", instance + 1, "of", str(instances) + "."
# Make initial Options.
for num_options in x_axis:
options, _ = make_near_optimal_phi_relative_options(
mdp,
state_abstr,
'eps-greedy',
num_rand_opts=num_options - 1,
eps=epsilon)
action_abstr = ActionAbstraction(
options=options, prim_actions=mdp.get_actions())
# Make agent.
AgentClass = RMaxAgent # DoubleQAgent, QLearningAgent, SarsaAgent
sa_aa_agent = AbstractionWrapper(
AgentClass,
agent_params={"actions": mdp.get_actions()},
state_abstr=state_abstr,
action_abstr=action_abstr,
name_ext="-$\\phi,O$")
_, _, value_per_episode = run_single_agent_on_mdp(
sa_aa_agent, mdp, episodes=episodes, steps=steps)
mdp.reset()
num_options_performance[(steps, episodes)][num_options].append(
value_per_episode[-1])
############
# Other types
# Just state abstraction.
steps, episodes = d_var_range[-1][0], d_var_range[-1][1]
sa_agent = AbstractionWrapper(AgentClass,
agent_params={"actions": mdp.get_actions()},
state_abstr=state_abstr,
action_abstr=None,
name_ext="-$\\phi$")
_, _, value_per_episode = run_single_agent_on_mdp(sa_agent,
mdp,
episodes=episodes,
steps=steps)
num_options_performance[(d_var_range[-1][0],
d_var_range[-1][1])]["phi"].append(
value_per_episode[-1])
y_axis["phi"] = [value_per_episode[-1]]
# Run random options.
options = make_fixed_random_options(mdp, state_abstr)
action_abstr = ActionAbstraction(options=options,
prim_actions=mdp.get_actions())
AgentClass = QLearningAgent
rand_opt_agent = AbstractionWrapper(
AgentClass,
agent_params={"actions": mdp.get_actions()},
state_abstr=state_abstr,
action_abstr=action_abstr,
name_ext="-$\\phi,O_{\text{random}}$")
_, _, value_per_episode = run_single_agent_on_mdp(rand_opt_agent,
mdp,
episodes=episodes,
steps=steps)
num_options_performance[(d_var_range[-1][0],
d_var_range[-1][1])]["random"].append(
value_per_episode[-1])
y_axis["random"] = [value_per_episode[-1]]
# Makeoptimal agent.
value_iter = ValueIteration(mdp)
value_iter.run_vi()
optimal_agent = FixedPolicyAgent(value_iter.policy)
_, _, value_per_episode = run_single_agent_on_mdp(optimal_agent,
mdp,
episodes=episodes,
steps=steps)
y_axis["optimal"] = [value_per_episode[-1]]
total_steps = d_var_range[0][0] * d_var_range[0][1]
# Confidence intervals.
for dependent_var in d_var_range:
for num_options in x_axis:
# Compute mean and standard error.
avg_for_n = float(
sum(num_options_performance[dependent_var]
[num_options])) / instances
std_deviation = np.std(
num_options_performance[dependent_var][num_options])
std_error = 1.96 * (std_deviation / math.sqrt(
len(num_options_performance[dependent_var][num_options])))
y_axis[dependent_var].append(avg_for_n)
conf_intervals[dependent_var].append(std_error)
plt.xlabel("$|O_\\phi|$")
plt.xlim([1, len(x_axis)])
plt.ylabel("$V^{\hat{\pi}_{O_\\phi}}(s_0)$")
plt.tight_layout() # Keeps the spacing nice.
# Add just state abstraction.
ep_val_del_q_phi = y_axis["phi"]
label = "$O_{\\phi}$" #" N=1e" + str(str(total_steps).count("0")) + "$"
plt.plot(x_axis, [ep_val_del_q_phi] * len(x_axis),
marker="+",
linestyle="--",
linewidth=1.0,
color=PLOT_COLORS[-1],
label=label)
# Add random options.
ep_val_del_q = y_axis["random"]
label = "$O_{random}$" #" N=1e" + str(str(total_steps).count("0")) + "$"
plt.plot(x_axis, [ep_val_del_q] * len(x_axis),
marker="x",
linestyle="--",
linewidth=1.0,
color=PLOT_COLORS[0]) #, label=label)
# Add optimal.
ep_val_optimal = y_axis["optimal"]
plt.plot(x_axis, [ep_val_optimal] * len(x_axis),
linestyle="-",
linewidth=1.0,
color=PLOT_COLORS[1]) #, label="$\\pi^*$")
for i, dependent_var in enumerate(d_var_range):
total_steps = dependent_var[0] * dependent_var[1]
label = "$O_{\\phi,Q_\\varepsilon^*}, N=1e" + str(
str(total_steps).count("0")) + "$"
plt.plot(x_axis,
y_axis[dependent_var],
marker="x",
color=PLOT_COLORS[i + 2],
linewidth=1.5,
label=label)
# Confidence intervals.
top = np.add(y_axis[dependent_var], conf_intervals[dependent_var])
bot = np.subtract(y_axis[dependent_var], conf_intervals[dependent_var])
plt.fill_between(x_axis,
top,
bot,
alpha=0.25,
color=PLOT_COLORS[i + 2])
plt.legend()
plt.savefig("branching_factor_results.pdf", format="pdf")
plt.cla()
plt.close()
def info_sa_planning_experiment(min_grid_size=5, max_grid_size=11, beta=10.0):
'''
Args:
min_grid_size (int)
max_grid_size (int)
beta (float): Hyperparameter for InfoSA.
Summary:
Writes num iterations and time (seconds) for planning with and without abstractions.
'''
vanilla_file = "vi.csv"
sa_file = "vi-$\\phi$.csv"
file_prefix = os.path.join("results", "planning-four_room")
clear_files(dir_name=file_prefix)
for grid_dim in xrange(min_grid_size, max_grid_size + 1):
# ======================
# == Make Environment ==
# ======================
mdp = FourRoomMDP(width=grid_dim, height=grid_dim, init_loc=(1, 1), goal_locs=[(grid_dim, grid_dim)], gamma=0.9)
# Get demo policy.
vi = ValueIteration(mdp)
vi.run_vi()
demo_policy = get_lambda_policy(make_det_policy_eps_greedy(vi.policy, vi.get_states(), mdp.get_actions(), epsilon=0.2))
# =======================
# == Make Abstractions ==
# =======================
pmf_s_phi, phi_pmf, abstr_policy = run_info_sa(mdp, demo_policy, iters=500, beta=beta, convergence_threshold=0.00001)
lambda_abstr_policy = get_lambda_policy(abstr_policy)
prob_s_phi = ProbStateAbstraction(phi_pmf)
crisp_s_phi = convert_prob_sa_to_sa(prob_s_phi)
# ============
# == Run VI ==
# ============
vanilla_vi = ValueIteration(mdp, delta=0.0001, sample_rate=25)
sa_vi = AbstractValueIteration(ground_mdp=mdp, state_abstr=crisp_s_phi, delta=0.0001, vi_sample_rate=25, amdp_sample_rate=25)
# ==========
# == Plan ==
# ==========
print "Running VIs."
start_time = time.clock()
vanilla_iters, vanilla_val = vanilla_vi.run_vi()
vanilla_time = round(time.clock() - start_time, 2)
mdp.reset()
start_time = time.clock()
sa_iters, sa_abs_val = sa_vi.run_vi()
sa_time = round(time.clock() - start_time, 2)
sa_val = evaluate_agent(FixedPolicyAgent(sa_vi.policy), mdp, instances=25)
print "\n" + "*"*20
print "Vanilla", "\n\t Iters:", vanilla_iters, "\n\t Value:", round(vanilla_val, 4), "\n\t Time:", vanilla_time
print
print "Phi:", "\n\t Iters:", sa_iters, "\n\t Value:", round(sa_val, 4), "\n\t Time:", sa_time
print "*"*20 + "\n\n"
write_datum(os.path.join(file_prefix, "iters", vanilla_file), vanilla_iters)
write_datum(os.path.join(file_prefix, "iters", sa_file), sa_iters)
write_datum(os.path.join(file_prefix, "times", vanilla_file), vanilla_time)
write_datum(os.path.join(file_prefix, "times", sa_file), sa_time)
def main():
# Setup MDP, Agents.
mdp = FourRoomMDP(11, 11, goal_locs=[(11, 11)], gamma=0.9, step_cost=0.0)
ql_agent = QLearningAgent(mdp.get_actions(), epsilon=0.2, alpha=0.4)
viz = parse_args()
# Choose viz type.
viz = "learning"
if viz == "value":
# Run experiment and make plot.
mdp.visualize_value()
elif viz == "policy":
# Viz policy
value_iter = ValueIteration(mdp)
value_iter.run_vi()
policy = value_iter.policy
mdp.visualize_policy(policy)
elif viz == "agent":
# Solve problem and show agent interaction.
print("\n", str(ql_agent), "interacting with", str(mdp))
run_single_agent_on_mdp(ql_agent, mdp, episodes=500, steps=200)
mdp.visualize_agent(ql_agent)
elif viz == "learning":
# Run experiment and make plot.
mdp.visualize_learning(ql_agent)
elif viz == "interactive":
mdp.visualize_interaction()
def run_learning_experiment():
"""
Summary:
Builds different sets of options and contrasts how RL algorithms
perform when learning with them.
"""
# Define MDP.
width, height = 11, 11
mdp = FourRoomMDP(width=width,
height=height,
goal_locs=[(width, height)],
slip_prob=0.05)
actions = mdp.get_actions()
# Make State Abstraction.
states, _ = ah.compute_reachable_state_space(mdp, sample_rate=50)
if isinstance(mdp, FourRoomMDP):
predicate = four_rooms_predicate_11x11
else:
predicate = reachable_in_n_steps_predicate
state_abstr = core.compute_phi_given_m(mdp,
predicate,
level=1,
states=states)
# Make initial Options.
num_rand_opts_to_add = 2
options, _ = make_near_optimal_phi_relative_options(
mdp,
state_abstr,
'eps-greedy',
num_rand_opts=num_rand_opts_to_add,
eps=0.05)
action_abstr = ActionAbstraction(options=options, prim_actions=actions)
action_abstr_w_prims = ActionAbstraction(options=options,
prim_actions=actions,
incl_primitives=True)
# Find eigen options.
# num_eigen_options = max(1, num_rand_opts_to_add - 1)
# eigen_options_init_all = find_eigenoptions(mdp, num_options=num_eigen_options, init_everywhere=True)
# eigen_options_w_prims = find_eigenoptions(mdp, num_options=num_eigen_options, init_everywhere=False)
# eigen_aa_init_all = ActionAbstraction(options=eigen_options_init_all, prim_actions=actions, incl_primitives=False)
# eigen_aa_w_prims = ActionAbstraction(options=eigen_options_w_prims, prim_actions=actions, incl_primitives=True)
# Make agent.
AgentClass = QLearningAgent #QLearningAgent #DoubleQAgent #DelayedQAgent
ql_agent = AgentClass(mdp.get_actions())
sa_aa_agent = AbstractionWrapper(AgentClass,
agent_params={"actions": actions},
state_abstr=state_abstr,
action_abstr=action_abstr_w_prims,
name_ext="-$\\phi,O$")
aa_agent = AbstractionWrapper(AgentClass,
agent_params={"actions": actions},
state_abstr=None,
action_abstr=action_abstr_w_prims,
name_ext="-$O$")
# aa_agent = AbstractionWrapper(AgentClass, agent_params={"actions": actions}, state_abstr=None, action_abstr=action_abstr_w_prims, name_ext="-$\\phi$")
# Eigen agents.
# eigen_agent_init_all = AbstractionWrapper(AgentClass, agent_params={"actions": actions}, state_abstr=None, action_abstr=eigen_aa_init_all, name_ext="-eigen_all")
# eigen_agent_w_prims = AbstractionWrapper(AgentClass, agent_params={"actions": actions}, state_abstr=None, action_abstr=eigen_aa_w_prims, name_ext="-eigen_w_prims")
agents = [ql_agent, aa_agent,
sa_aa_agent] #, eigen_agent_init_all, eigen_agent_w_prims]
# Run.
if isinstance(mdp, FourRoomMDP):
run_agents_on_mdp(agents, mdp, instances=10, episodes=500, steps=50)
else:
run_agents_on_mdp(agents, mdp, instances=10, episodes=100, steps=10)
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)
s = torch.tensor((x, y)).float()
action = torch.argmax(
alg.action_distr(s, torch.tensor(skill)).logits).item()
action_word = dict(enumerate(alg.env.action_map)).get(
action, 'term')
policy[x, y] = action_word
return policy
plt.figure()
#%%
reset_seeds(0)
from simple_rl.tasks import FourRoomMDP
from notebooks.simple_rl_env import SimpleGymEnv
env = SimpleGymEnv(FourRoomMDP(12, 12, goal_locs=[(12, 12)]))
env.render()
s0 = torch.as_tensor(env.reset(), dtype=torch.float32)
ndim_s = len(env.observation_space)
n_actions = env.action_space.n
n_skills = 40
gamma = 0.99
max_steps_per_skill = 10
n_units = 32
lr = 1e-3
alg = VIC
# alg = DIAYN
alg = alg(env, ndim_s, n_actions, n_skills, gamma, max_steps_per_skill,
n_units, lr)
start_policies = [get_policy(alg, env.mdp, s) for s in range(n_skills)]
#%%
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 main():
# Setup MDP, Agents.
mdp = FourRoomMDP(5, 5, goal_locs=[(5, 5)], gamma=0.99, step_cost=0.01)
# mdp = make_grid_world_from_file("octogrid.txt", num_goals=12, randomize=False)
ql_agent = QLearningAgent(mdp.get_actions(), epsilon=0.2, alpha=0.5)
rm_agent = RMaxAgent(mdp.get_actions())
viz = parse_args()
viz = "learning"
if viz == "value":
# Run experiment and make plot.
mdp.visualize_value()
elif viz == "policy":
# Viz policy
value_iter = ValueIteration(mdp)
value_iter.run_vi()
policy = value_iter.policy
mdp.visualize_policy(policy)
elif viz == "agent":
# Solve problem and show agent interaction.
print("\n", str(ql_agent), "interacting with", str(mdp))
run_single_agent_on_mdp(ql_agent, mdp, episodes=500, steps=200)
mdp.visualize_agent(ql_agent)
elif viz == "learning":
# Run experiment and make plot.
mdp.visualize_learning(ql_agent)
pass parent_dir = os.path.abspath(os.path.join(os.getcwd(), os.pardir)) sys.path.insert(0, parent_dir) from simple_rl.agents import QLearningAgent, RandomAgent, RMaxAgent from simple_rl.planning import ValueIteration from simple_rl.tasks import GridWorldMDP from simple_rl.run_experiments import run_agents_on_mdp # # Setup MDP. # mdp = GridWorldMDP(width=6, height=6, init_loc=(1, 1), goal_locs=[(6, 6)]) # # # Setup Agents. # ql_agent = QLearningAgent(actions=mdp.get_actions()) # rand_agent = RandomAgent(actions=mdp.get_actions()) # rmax_agent = RMaxAgent(actions=mdp.get_actions(), horizon=3, s_a_threshold=1) # # # Run experiment and make plot. # run_agents_on_mdp([ql_agent, rand_agent, rmax_agent], mdp, instances=5, episodes=100, steps=40, reset_at_terminal=True, # verbose=False) from simple_rl.tasks import FourRoomMDP from simple_rl.tasks.grid_world import grid_visualizer four_room_mdp = FourRoomMDP(9, 9, goal_locs=[(9, 9), (5, 4)], gamma=0.95) # Run experiment and make plot. # four_room_mdp.visualize_value() four_room_mdp.visualize_interaction() # four_room_mdp.visualize_policy()
