def build_point_option_agent(mdp,
pairs,
agent=QLearningAgent,
policy='vi',
name='-abstr'):
# pairs should be a List of pair.
# Pair is conposed of two lists.
# Each list has init/term states.
goal_based_options = aa_helpers.make_point_options(mdp,
pairs,
policy=policy)
goal_based_aa = ActionAbstraction(prim_actions=mdp.get_actions(),
options=goal_based_options,
use_prims=True)
# num_feats = mdp.get_num_state_feats()
option_agent = AbstractionWrapper(
agent,
agent_params={"actions": mdp.get_actions()},
action_abstr=goal_based_aa,
name_ext=name)
# option_agent = AbstractionWrapper(LinearQAgent, agent_params={"actions":mdp.get_actions(), "num_features":num_feats}, action_abstr=goal_based_aa, name_ext=name)
# option_agent = AbstractionWrapper(QLearningAgent, agent_params={"actions":mdp.get_actions()}, action_abstr=goal_based_aa, name_ext=name)
return option_agent
def build_subgoal_option_agent(mdp,
subgoals,
init_region,
agent=QLearningAgent,
vectors=None,
name='-abstr',
n_trajs=50,
n_steps=100,
classifier='list',
policy='vi'):
# print('sbugoals=', subgoals)
goal_based_options = aa_helpers.make_subgoal_options(mdp,
subgoals,
init_region,
vectors=vectors,
n_trajs=n_trajs,
n_steps=n_steps,
classifier=classifier,
policy=policy)
goal_based_aa = ActionAbstraction(prim_actions=mdp.get_actions(),
options=goal_based_options,
use_prims=True)
# num_feats = mdp.get_num_state_feats()
option_agent = AbstractionWrapper(
agent,
agent_params={"actions": mdp.get_actions()},
action_abstr=goal_based_aa,
name_ext=name)
return option_agent
def main(open_plot=True):
# Setup MDP, Agents.
mdp = GridWorldMDP(width=10, height=10, init_loc=(1, 1), goal_locs=[(10, 10)])
ql_agent = QLearningAgent(actions=mdp.get_actions())
rand_agent = RandomAgent(actions=mdp.get_actions())
abstr_identity_agent = AbstractionWrapper(QLearningAgent, agent_params={"epsilon":0.9}, actions=mdp.get_actions())
# Run experiment and make plot.
run_agents_on_mdp([ql_agent, rand_agent, abstr_identity_agent], mdp, instances=5, episodes=100, steps=150, open_plot=open_plot)
def main(open_plot=True):
# Setup MDP, Agents.
mdp_distr = make_mdp.make_mdp_distr(mdp_class="four_room")
ql_agent = QLearningAgent(actions=mdp_distr.get_actions())
rand_agent = RandomAgent(actions=mdp_distr.get_actions())
# Make goal-based option agent.
goal_based_options = aa_helpers.make_goal_based_options(mdp_distr)
goal_based_aa = ActionAbstraction(prim_actions=mdp_distr.get_actions(),
options=goal_based_options)
option_agent = AbstractionWrapper(QLearningAgent,
actions=mdp_distr.get_actions(),
action_abstr=goal_based_aa)
# Run experiment and make plot.
run_agents_lifelong([ql_agent, rand_agent, option_agent],
mdp_distr,
samples=10,
episodes=100,
steps=150,
open_plot=open_plot)
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 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)