def main(open_plot=True): # Setup MDP. mdp = GridWorldMDP(width=4, height=3, init_loc=(1, 1), goal_locs=[(4, 3)], lava_locs=[(4, 2)], gamma=0.95, walls=[(2, 2)], slip_prob=0.05) # Make agents. ql_agent = QLearningAgent(actions=mdp.get_actions()) rand_agent = RandomAgent(actions=mdp.get_actions()) tabular_agent = CherryQAgent(mdp, model=lambda *x: ActionValueFunction(*x, init=1.0), name='Tabular', lr=0.7) linear_agent = CherryQAgent(mdp, model=lambda *x: nn.Linear(*x), name='Linear', lr=0.1) mlp_agent = CherryQAgent(mdp, model=lambda *x: MLP(*x), name='MLP', lr=0.07) # Run experiment and make plot. agents = [rand_agent, ql_agent, tabular_agent, linear_agent, mlp_agent] run_agents_on_mdp(agents, mdp, instances=10, episodes=50, steps=50, open_plot=open_plot)
def main(): # Command line args. task, rom = parse_args() # Setup the MDP. mdp = choose_mdp(task, rom) actions = mdp.get_actions() gamma = mdp.get_gamma() # Setup agents. from simple_rl.agents import RandomAgent, QLearningAgent random_agent = RandomAgent(actions) qlearner_agent = QLearningAgent(actions, gamma=gamma, explore="uniform") agents = [qlearner_agent, random_agent] # Run Agents. if isinstance(mdp, MarkovGameMDP): # Markov Game. agents = { qlearner_agent.name: qlearner_agent, random_agent.name: random_agent } play_markov_game(agents, mdp, instances=100, episodes=1, steps=500) else: # Regular experiment. run_agents_on_mdp(agents, mdp, instances=50, episodes=1, steps=2000)
def main(open_plot=True): # gym_mdp = GridWorldMDP(width=10, height=10, init_loc=(1,1), goal_locs=[(10,10)]) # num_feats = gym_mdp.get_num_state_feats() # lin_agent = QLearnerAgent(gym_mdp.actions, alpha=0.4, epsilon=0.4) # rand_agent = RandomAgent(gym_mdp.actions) # run_agents_on_mdp([lin_agent, rand_agent], gym_mdp, instances=50, episodes=200, steps=100, open_plot=open_plot) # gym_mdp = GridWorldMDP(width=10, height=10, init_loc=(1,1), goal_locs=[(10,10)]) # num_feats = gym_mdp.get_num_state_feats() # lin_agent = LinearQLearnerAgent(gym_mdp.actions, num_features=num_feats, alpha=0.4, epsilon=0.4, anneal=False,rbf=True) # rand_agent = RandomAgent(gym_mdp.actions) # run_agents_on_mdp([lin_agent, rand_agent], gym_mdp, instances=50, episodes=200, steps=100, open_plot=open_plot,verbose=True) gym_mdp = GymMDP(env_name='CartPole-v0', render=False) num_feats = gym_mdp.get_num_state_feats() lin_agent = LinearQLearnerAgent(gym_mdp.actions, num_features=num_feats, alpha=0.4, epsilon=0.4, anneal=False, rbf=True) rand_agent = RandomAgent(gym_mdp.actions) run_agents_on_mdp([lin_agent, rand_agent], gym_mdp, instances=5, episodes=1000, steps=100, open_plot=open_plot)
def main(): # Setup MDP, Agents. size = 5 agent = { "x": 1, "y": 1, "dx": 1, "dy": 0, "dest_x": size, "dest_y": size, "has_block": 0 } blocks = [{"x": size, "y": 1}] lavas = [{ "x": x, "y": y } for x, y in map(lambda z: (z + 1, (size + 1) / 2), range(size))] mdp = TrenchOOMDP(size, size, agent, blocks, lavas) 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=30, episodes=250, steps=250)
def main(open_plot=True): # Setup MDP. mdp = GridWorldMDP(width=8, height=3, init_loc=(1, 1), goal_locs=[(8, 3)], lava_locs=[(4, 2)], gamma=0.95, walls=[(2, 2)], slip_prob=0.05) # Make agents. 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=20, episodes=300, steps=20, open_plot=open_plot, track_success=True, success_reward=1)
def main(open_plot=True): # Taxi initial state attributes.. agent = {"x": 1, "y": 1, "has_passenger": 0} passengers = [{"x": 3, "y": 2, "dest_x": 2, "dest_y": 3, "in_taxi": 0}] walls = [] mdp = TaxiOOMDP(width=4, height=4, agent=agent, walls=walls, passengers=passengers) # Agents. ql_agent = QLearningAgent(actions=mdp.get_actions()) rand_agent = RandomAgent(actions=mdp.get_actions()) viz = False if viz: # Visualize Taxi. run_single_agent_on_mdp(ql_agent, mdp, episodes=50, steps=1000) mdp.visualize_agent(ql_agent) else: # Run experiment and make plot. run_agents_on_mdp([ql_agent, rand_agent], mdp, instances=10, episodes=1, steps=500, reset_at_terminal=True, open_plot=open_plot)
def main(): # Setup MDP, Agents. size = 5 agent = { "x": 1, "y": 1, "dx": 1, "dy": 0, "dest_x": size, "dest_y": size, "has_block": 0 } blocks = [{"x": size, "y": 1}] lavas = [{ "x": x, "y": y } for x, y in map(lambda z: (z + 1, (size + 1) / 2), xrange(size))] mdp = TrenchOOMDP(size, size, agent, blocks, lavas) ql_agent = QLearnerAgent(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=30, episodes=250, steps=250) vi = ValueIteration(mdp, delta=0.0001, max_iterations=5000) iters, val = vi.run_vi() print " done." states = vi.get_states() num_states = len(states) print num_states, states
def main(open_plot=True): # Make MDP distribution, agents. mdp_distr = make_mdp.make_mdp_distr(mdp_class="four_room") ql_agent = QLearnerAgent(actions=mdp_distr.get_actions()) rand_agent = RandomAgent(actions=mdp_distr.get_actions()) # Run experiment and make plot. run_agents_multi_task([ql_agent, rand_agent], mdp_distr, task_samples=50, episodes=1, steps=1500, reset_at_terminal=True, open_plot=open_plot)
def main(open_plot=True): # Gym MDP gym_mdp = GymMDP(env_name='CartPole-v0', render=False) num_feats = gym_mdp.get_num_state_feats() # Setup agents and run. lin_agent = LinearQLearnerAgent(gym_mdp.actions, num_features=num_feats, alpha=0.4, epsilon=0.4, anneal=True) rand_agent = RandomAgent(gym_mdp.actions) run_agents_on_mdp([lin_agent, rand_agent], gym_mdp, instances=10, episodes=30, steps=10000, open_plot=open_plot)
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(): # Setup MDP. actual_args = { "width": 10, "height": 10, "init_loc": (1, 1), "goal_locs": [(10, 10)], "lava_locs": [(1, 10), (3, 10), (5, 10), (7, 10), (9, 10)], "gamma": 0.9, "walls": [ (2, 2), (2, 3), (2, 4), (2, 5), (2, 6), (2, 7), (2, 8), (2, 9), (4, 2), (4, 3), (4, 4), (4, 5), (4, 6), (4, 7), (4, 8), (4, 9), (6, 2), (6, 3), (6, 4), (6, 5), (6, 6), (6, 7), (6, 8), (6, 9), (8, 2), (8, 3), (8, 4), (8, 5), (8, 6), (8, 7), (8, 8), (8, 9) ], "slip_prob": 0.01, "lava_cost": 1.0, "step_cost": 0.1 } mdp = GridWorldMDP(**actual_args) # Initialize the custom Q function for a q-learning agent. This should be equivalent to potential shaping. # This should cause the Q agent to learn more quickly. custom_q = defaultdict(lambda: defaultdict(lambda: 0)) custom_q[GridWorldState(5, 1)]['right'] = 1.0 custom_q[GridWorldState(2, 1)]['right'] = 1.0 # Make a normal q-learning agent and another initialized with the custom_q above. # Finally, make a random agent to compare against. ql_agent = QLearningAgent(actions=mdp.get_actions(), epsilon=0.2, alpha=0.4) ql_agent_pot = QLearningAgent(actions=mdp.get_actions(), epsilon=0.2, alpha=0.4, custom_q_init=custom_q, name="PotQ") rand_agent = RandomAgent(actions=mdp.get_actions()) # Run experiment and make plot. run_agents_on_mdp([ql_agent, ql_agent_pot, rand_agent], mdp, instances=2, episodes=60, steps=200, open_plot=True, verbose=True)
def main(open_plot=True): state_colors = defaultdict(lambda:defaultdict(lambda:"white")) state_colors[3][2] = "red" # Setup MDP, Agents. mdp = ColoredGridWorldMDP(state_colors) ql_agent = QLearnerAgent(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=15, episodes=500, steps=40, open_plot=open_plot)
def collect_dataset(mdp, samples=10000, learning_agent=None): ''' Args: mdp (simple_rl.MDP) samples (int) learning_agent (simple_rl.Agent): If None, a random agent is used. Otherwise collects data based on its learning. Returns: (set) ''' if learning_agent is None: learning_agent = RandomAgent(mdp.get_actions()) cur_state = mdp.get_init_state() reward = 0 visited_states = set([cur_state]) # Set initial state params. init_state_params = {} last_x = 0 + np.random.randn(1)[0] init_state_params["x"] = last_x init_state_params["x_dot"] = 0 init_state_params["theta"] = 0 init_state_params["theta_dot"] = 0 for i in range(samples): action = learning_agent.act(cur_state, reward) reward, next_state = mdp.execute_agent_action(action) visited_states.add(next_state) if next_state.is_terminal(): init_state_params["x"] = np.random.randn(1)[0] mdp.reset(init_state_params) learning_agent.end_of_episode() cur_state = mdp.get_init_state() reward = 0 else: cur_state = next_state return visited_states
def _setup_agents(solar_mdp): ''' Args: solar_mdp (SolarOOMDP) Returns: (list): of Agents ''' # Get relevant MDP params. actions, gamma, panel_step = solar_mdp.get_actions(), solar_mdp.get_gamma( ), solar_mdp.get_panel_step() # Setup fixed agent. static_agent = FixedPolicyAgent(tb.static_policy, name="fixed-panel") optimal_agent = FixedPolicyAgent(tb.optimal_policy, name="optimal") # Grena single axis and double axis trackers from time/loc. grena_tracker = SolarTracker(tb.grena_tracker, panel_step=panel_step, dual_axis=True) grena_tracker_agent = FixedPolicyAgent(grena_tracker.get_policy(), name="grena-tracker") # Setup RL agents alpha, epsilon = 0.3, 0.3 num_features = solar_mdp.get_num_state_feats() lin_ucb_agent = LinUCBAgent(actions, name="lin-ucb", alpha=0.3) #, alpha=0.2) ql_lin_approx_agent_g0 = LinearQLearnerAgent(actions, num_features=num_features, name="ql-lin-g0", alpha=alpha, epsilon=epsilon, gamma=0, rbf=True, anneal=True) ql_lin_approx_agent = LinearQLearnerAgent(actions, num_features=num_features, name="ql-lin", alpha=alpha, epsilon=epsilon, gamma=gamma, rbf=True, anneal=True) # sarsa_lin_rbf_agent = LinearApproxSarsaAgent(actions, name="sarsa-lin", alpha=alpha, epsilon=epsilon, gamma=gamma, rbf=True, anneal=False) random_agent = RandomAgent(actions) # Regular experiments. agents = [ ql_lin_approx_agent, lin_ucb_agent, grena_tracker_agent, static_agent ] return agents
def main(): import OptimalBeliefAgentClass # Setup multitask setting. # R ~ D : Puddle, Rock Sample # G ~ D : octo, four_room # T ~ D : grid mdp_class, is_goal_terminal, samples = parse_args() mdp_distr = make_mdp_distr(mdp_class=mdp_class, is_goal_terminal=is_goal_terminal) mdp_distr.set_gamma(0.99) actions = mdp_distr.get_actions() # Compute average MDP. print "Making and solving avg MDP...", sys.stdout.flush() avg_mdp = compute_avg_mdp(mdp_distr) avg_mdp_vi = ValueIteration(avg_mdp, delta=0.001, max_iterations=1000, sample_rate=5) iters, value = avg_mdp_vi.run_vi() print "done." #, iters, value sys.stdout.flush() # Agents. print "Making agents...", sys.stdout.flush() mdp_distr_copy = copy.deepcopy(mdp_distr) opt_stoch_policy = compute_optimal_stoch_policy(mdp_distr_copy) opt_stoch_policy_agent = FixedPolicyAgent(opt_stoch_policy, name="$\pi_{prior}$") opt_belief_agent = OptimalBeliefAgentClass.OptimalBeliefAgent( mdp_distr, actions) vi_agent = FixedPolicyAgent(avg_mdp_vi.policy, name="$\pi_{avg}$") rand_agent = RandomAgent(actions, name="$\pi^u$") ql_agent = QLearningAgent(actions) print "done." agents = [vi_agent, opt_stoch_policy_agent, rand_agent, opt_belief_agent] # Run task. run_agents_multi_task(agents, mdp_distr, task_samples=samples, episodes=1, steps=100, reset_at_terminal=False, track_disc_reward=False, cumulative_plot=True)
def main(): # create mdp using own definition mdp = tfeMDP() # Three different agents to compare how each do against each other rand_agent = RandomAgent(actions=mdp.get_actions()) rmax_agent = RMaxAgent(actions=mdp.get_actions()) agent = QLearningAgent(actions=mdp.get_actions()) # Function that actually runs everything and generates the appropriate # graphs and statistics defining how each agent did run_agents_on_mdp([agent, rmax_agent, rand_agent], mdp, instances=200, episodes=100, steps=1000)
def main(): # Make MDP distribution, agents. mdp_distr = make_mdp_distr(mdp_class="grid") ql_agent = QLearnerAgent(actions=mdp_distr.get_actions()) rand_agent = RandomAgent(actions=mdp_distr.get_actions()) # Run experiment and make plot. run_agents_multi_task([ql_agent, rand_agent], mdp_distr, task_samples=30, episodes=100, steps=50, reset_at_terminal=True, include_optimal=True)
def main(open_plot=True): # Setup MDP, Agents. mdp = BanditMDP() lin_agent = LinUCBAgent(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, lin_agent, rand_agent], mdp, instances=10, episodes=1, steps=500, open_plot=open_plot)
def main(open_plot=True): # Gym MDP gym_mdp = GymMDP(env_name='Breakout-v0', render=False) num_feats = gym_mdp.get_num_state_feats() # Setup agents and run. rand_agent = RandomAgent(gym_mdp.get_actions()) lin_q_agent = LinearQAgent(gym_mdp.get_actions(), num_feats) run_agents_on_mdp([lin_q_agent, rand_agent], gym_mdp, instances=5, episodes=50000, steps=200, open_plot=open_plot, verbose=False)
def main(open_plot=True): # Setup MDP, Agents. mdp = GridWorldMDP(width=10, height=10, init_loc=(1, 1), goal_locs=[(10, 10)]) ql_agent = QLearnerAgent(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=100, steps=150, open_plot=open_plot)
def _setup_agents(solar_mdp): ''' Args: solar_mdp (SolarOOMDP) Returns: (list): of Agents ''' # Get relevant MDP params. actions, gamma, panel_step = solar_mdp.get_actions(), solar_mdp.get_gamma( ), solar_mdp.get_panel_step() # Setup fixed agent. static_agent = FixedPolicyAgent(tb.static_policy, name="fixed-panel") optimal_agent = FixedPolicyAgent(tb.optimal_policy, name="optimal") # Grena single axis and double axis trackers from time/loc. grena_tracker = SolarTracker(tb.grena_tracker, panel_step=panel_step, dual_axis=solar_mdp.dual_axis, actions=solar_mdp.get_bandit_actions()) grena_tracker_agent = FixedPolicyAgent(grena_tracker.get_policy(), name="grena-tracker") # Setup RL agents alpha, epsilon = 0.1, 0.05 rand_init = True num_features = solar_mdp.get_num_state_feats() lin_ucb_agent = LinUCBAgent(solar_mdp.get_bandit_actions(), context_size=num_features, name="lin-ucb", rand_init=rand_init, alpha=2.0) # sarsa_agent_g0 = LinearSarsaAgent(actions, num_features=num_features, name="sarsa-lin-g0", rand_init=rand_init, alpha=alpha, epsilon=epsilon, gamma=0, rbf=False, anneal=True) # sarsa_agent = LinearSarsaAgent(actions, num_features=num_features, name="sarsa-lin", rand_init=rand_init, alpha=alpha, epsilon=epsilon, gamma=gamma, rbf=False, anneal=True) ql_agent = QLearningAgent(actions, alpha=alpha, epsilon=epsilon, gamma=gamma) random_agent = RandomAgent(actions) # Regular experiments. # agents = [lin_ucb_agent, sarsa_agent, sarsa_agent_g0, grena_tracker_agent, static_agent] agents = [grena_tracker_agent, static_agent] return agents
def main(): # ======================== # === Make Environment === # ======================== mdp_class = "four_room" gamma = 1.0 environment = make_mdp.make_mdp_distr(mdp_class=mdp_class, step_cost=0.01, grid_dim=15, gamma=gamma) actions = environment.get_actions() # ========================== # === Make SA, AA Stacks === # ========================== sa_stack, aa_stack = hierarchy_helpers.make_hierarchy(environment, num_levels=2) # Debug. print "\n" + ("=" * 30) + "\n== Done making abstraction. ==\n" + ("=" * 30) + "\n" sa_stack.print_state_space_sizes() aa_stack.print_action_spaces_sizes() # =================== # === Make Agents === # =================== # baseline_agent = QLearnerAgent(actions) agent_class = QLearnerAgent baseline_agent = agent_class(actions, gamma=gamma) rand_agent = RandomAgent(actions) # hierarch_r_max = HRMaxAgent(actions, sa_stack=sa_stack, aa_stack=aa_stack) l0_hierarch_agent = HierarchyAgent(agent_class, sa_stack=sa_stack, aa_stack=aa_stack, cur_level=0, name_ext="-$l_0$") l1_hierarch_agent = HierarchyAgent(agent_class, sa_stack=sa_stack, aa_stack=aa_stack, cur_level=1, name_ext="-$l_1$") # l2_hierarch_agent = HierarchyAgent(agent_class, sa_stack=sa_stack, aa_stack=aa_stack, cur_level=2, name_ext="-$l_2$") dynamic_hierarch_agent = DynamicHierarchyAgent(agent_class, sa_stack=sa_stack, aa_stack=aa_stack, cur_level=1, name_ext="-$d$") # dynamic_rmax_hierarch_agent = DynamicHierarchyAgent(RMaxAgent, sa_stack=sa_stack, aa_stack=aa_stack, cur_level=1, name_ext="-$d$") print "\n" + ("=" * 26) print "== Running experiments. ==" print "=" * 26 + "\n" # ====================== # === Run Experiment === # ====================== agents = [dynamic_hierarch_agent, baseline_agent] run_agents_multi_task(agents, environment, task_samples=10, steps=20000, episodes=1, reset_at_terminal=True)
def main(): args = parse_args() mdp = generate_MDP(args.width, args.height, args.i_loc, args.g_loc, args.l_loc, args.gamma, args.Walls, args.slip) ql_agent = QLearningAgent(mdp.get_actions(), epsilon=args.epsilon, alpha=args.alpha, explore=args.explore, anneal=args.anneal) viz = args.mode if viz == "value": # --> Color corresponds to higher value. # Run experiment and make plot. mdp.visualize_value() elif viz == "policy": # Viz policy value_iter = ValueIteration(mdp) value_iter.run_vi() mdp.visualize_policy_values( (lambda state: value_iter.policy(state)), (lambda state: value_iter.value_func[state])) elif viz == "agent": # --> Press <spacebar> to advance the agent. # First let the agent solve the problem and then visualize the agent's resulting policy. print("\n", str(ql_agent), "interacting with", str(mdp)) rand_agent = RandomAgent(actions=mdp.get_actions()) run_agents_on_mdp([rand_agent, ql_agent], mdp, open_plot=True, episodes=60, steps=200, instances=5, success_reward=1) # mdp.visualize_agent(ql_agent) elif viz == "learning": # --> Press <r> to reset. # Show agent's interaction with the environment. mdp.visualize_learning(ql_agent, delay=0.005, num_ep=500, num_steps=200)
def main(): # ======================== # === Make Environment === # ======================== mdp_class = "hrooms" environment = make_mdp.make_mdp_distr(mdp_class=mdp_class) actions = environment.get_actions() # ========================== # === Make SA, AA Stacks === # ========================== # sa_stack, aa_stack = aa_stack_h.make_random_sa_diropt_aa_stack(environment, max_num_levels=3) sa_stack, aa_stack = hierarchy_helpers.make_hierarchy(environment, num_levels=3) # Debug. print("\n" + ("=" * 30)) print("== Done making abstraction. ==") print("=" * 30 + "\n") sa_stack.print_state_space_sizes() print("Num Action Abstractions:", len(aa_stack.get_aa_list())) # =================== # === Make Agents === # =================== baseline_agent = QLearningAgent(actions) rmax_agent = RMaxAgent(actions) rand_agent = RandomAgent(actions) l0_hierarch_agent = HierarchyAgent(QLearningAgent, sa_stack=sa_stack, aa_stack=aa_stack, cur_level=0, name_ext="-$l_0$") l1_hierarch_agent = HierarchyAgent(QLearningAgent, sa_stack=sa_stack, aa_stack=aa_stack, cur_level=1, name_ext="-$l_1$") # l2_hierarch_agent = HierarchyAgent(QLearningAgent, sa_stack=sa_stack, aa_stack=aa_stack, cur_level=2, name_ext="-$l_2$") dynamic_hierarch_agent = DynamicHierarchyAgent(QLearningAgent, sa_stack=sa_stack, aa_stack=aa_stack, cur_level=1, name_ext="-$d$") # dynamic_rmax_hierarch_agent = DynamicHierarchyAgent(RMaxAgent, sa_stack=sa_stack, aa_stack=aa_stack, cur_level=1, name_ext="-$d$") print("\n" + ("=" * 26)) print("== Running experiments. ==") print("=" * 26 + "\n") # ====================== # === Run Experiment === # ====================== agents = [l1_hierarch_agent, dynamic_hierarch_agent, baseline_agent] run_agents_multi_task(agents, environment, task_samples=10, steps=1500, episodes=1, reset_at_terminal=True)
def info_sa_compare_policies(mdp, demo_policy_lambda, beta=3.0, is_deterministic_ib=False, is_agent_in_control=False): ''' Args: mdp (simple_rl.MDP) demo_policy_lambda (lambda : simple_rl.State --> str) beta (float) is_deterministic_ib (bool): If True, run DIB, else IB. is_agent_in_control (bool): If True, runs the DIB in agent_in_control.py instead. Summary: Runs info_sa and compares the value of the found policy with the demonstrator policy. ''' if is_agent_in_control: # Run info_sa with the agent controlling the MDP. pmf_s_phi, phi_pmf, abstr_policy_pmf = agent_in_control.run_agent_in_control_info_sa(mdp, demo_policy_lambda, rounds=100, iters=500, beta=beta, is_deterministic_ib=is_deterministic_ib) else: # Run info_sa. pmf_s_phi, phi_pmf, abstr_policy_pmf = run_info_sa(mdp, demo_policy_lambda, iters=500, beta=beta, convergence_threshold=0.00001, is_deterministic_ib=is_deterministic_ib) # Make demonstrator agent and random agent. demo_agent = FixedPolicyAgent(demo_policy_lambda, name="$\\pi_d$") rand_agent = RandomAgent(mdp.get_actions(), name="$\\pi_u$") # Make abstract agent. lambda_abstr_policy = get_lambda_policy(abstr_policy_pmf) prob_s_phi = ProbStateAbstraction(phi_pmf) crisp_s_phi = convert_prob_sa_to_sa(prob_s_phi) abstr_agent = AbstractionWrapper(FixedPolicyAgent, state_abstr=crisp_s_phi, agent_params={"policy":lambda_abstr_policy, "name":"$\\pi_\\phi$"}, name_ext="") # Run. run_agents_on_mdp([demo_agent, abstr_agent, rand_agent], mdp, episodes=1, steps=1000) non_zero_abstr_states = [x for x in pmf_s_phi.values() if x > 0] # Print state space sizes. demo_vi = ValueIteration(mdp) print "\nState Spaces Sizes:" print "\t|S| =", demo_vi.get_num_states() print "\tH(S_\\phi) =", entropy(pmf_s_phi) print "\t|S_\\phi|_crisp =", crisp_s_phi.get_num_abstr_states() print "\tdelta_min =", min(non_zero_abstr_states) print "\tnum non zero states =", len(non_zero_abstr_states) print
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)]) 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=10, episodes=1, steps=20, open_plot=open_plot)
def main(): # Setup MDP. w = 6 h = 6 mdp = GridWorld(width=w, height=h, init_loc=(1, 1), goal_locs=[(6, 6)], slip_prob=.1) # Setup Agents. rand_agent = RandomAgent(actions=mdp.get_actions()) ql_agent = QLearningAgent(actions=mdp.get_actions()) # Compute number of samples for R-MAX to achieve epsilon optimal behavior with high probability (1 - delta) compute_n_samples = False if compute_n_samples: epsilon = .1 delta = .05 m_r = np.log(2. / delta) / (2. * epsilon**2) m_t = 2. * (np.log(2**(float(w * h)) - 2.) - np.log(delta)) / (epsilon **2) n_samples = int(max(m_r, m_t)) else: n_samples = 30 simple_rl_rmax_agent = RMaxAgent(actions=mdp.get_actions(), gamma=.9, horizon=3, s_a_threshold=n_samples, name='SimpleRL-R-MAX') rmax_agent = RMax(actions=mdp.get_actions(), gamma=.9, count_threshold=n_samples) # Run experiment and make plot. run_agents_on_mdp([rand_agent, ql_agent, rmax_agent, simple_rl_rmax_agent], mdp, instances=5, episodes=100, steps=20, reset_at_terminal=True, verbose=False)
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 main(open_plot=True): # Setup MDP. args = parse_args() mdp = generate_MDP(args.width, args.height, args.i_loc, args.g_loc, args.l_loc, args.gamma, args.Walls, args.slip) if args.visualize: value_iter = ValueIteration(mdp) value_iter.run_vi() mdp.visualize_policy_values( (lambda state: value_iter.policy(state)), (lambda state: value_iter.value_func[state])) else: custom_q = parse_custom_q_table(args.custom_q, args.default_q) agents = [] for agent in args.agents: if agent == 'q_learning': agents.append(QLearningAgent(actions=mdp.get_actions())) elif agent == 'potential_q': agents.append( QLearningAgent(actions=mdp.get_actions(), custom_q_init=custom_q, name="Potential_Q")) elif agent == 'random': agents.append(RandomAgent(actions=mdp.get_actions())) elif agent == 'rmax': agents.append(RMaxAgent(mdp.get_actions())) # Run experiment and make plot. run_agents_on_mdp(agents, mdp, instances=1, episodes=100, steps=100, open_plot=open_plot, verbose=True)
def main(open_plot=True): # Setup MDP. mdp = GridWorldMDP(width=4, height=3, init_loc=(1, 1), goal_locs=[(4, 3)], lava_locs=[(4, 2)], gamma=0.95, walls=[(2, 2)]) # Make agents. 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=50, steps=25, open_plot=open_plot, track_disc_reward=False)