def main(): parameters = parse_mover_problem_parameters() filename = 'pidx_%d_planner_seed_%d.pkl' % (parameters.pidx, parameters.planner_seed) save_dir = make_and_get_save_dir(parameters, filename) set_seed(parameters.pidx) problem_env = PaPMoverEnv(parameters.pidx) goal_object_names = [obj.GetName() for obj in problem_env.objects[:parameters.n_objs_pack]] goal_region_name = [problem_env.regions['home_region'].name] goal = goal_region_name + goal_object_names problem_env.set_goal(goal) goal_entities = goal_object_names + goal_region_name if parameters.use_shaped_reward: reward_function = ShapedRewardFunction(problem_env, goal_object_names, goal_region_name[0], parameters.planning_horizon) else: reward_function = GenericRewardFunction(problem_env, goal_object_names, goal_region_name[0], parameters.planning_horizon) motion_planner = OperatorBaseMotionPlanner(problem_env, 'prm') problem_env.set_reward_function(reward_function) problem_env.set_motion_planner(motion_planner) learned_q = None prior_q = None if parameters.use_learned_q: learned_q = load_learned_q(parameters, problem_env) v_fcn = lambda state: -len(get_objects_to_move(state, problem_env)) if parameters.planner == 'mcts': planner = MCTS(parameters, problem_env, goal_entities, prior_q, learned_q) elif parameters.planner == 'mcts_with_leaf_strategy': planner = MCTSWithLeafStrategy(parameters, problem_env, goal_entities, v_fcn, learned_q) else: raise NotImplementedError set_seed(parameters.planner_seed) search_time_to_reward, plan = planner.search(max_time=parameters.timelimit) pickle.dump({"search_time_to_reward": search_time_to_reward, 'plan': plan, 'n_nodes': len(planner.tree.get_discrete_nodes())}, open(save_dir+filename, 'wb'))
def main(): commit_hash = get_commit_hash() parameters = parse_mover_problem_parameters() filename = 'pidx_%d_planner_seed_%d.pkl' % (parameters.pidx, parameters.planner_seed) save_dir = make_and_get_save_dir(parameters, filename, commit_hash) solution_file_name = save_dir + filename is_problem_solved_before = os.path.isfile(solution_file_name) print solution_file_name if is_problem_solved_before and not parameters.f: print "***************Already solved********************" with open(solution_file_name, 'rb') as f: trajectory = pickle.load(f) tottime = trajectory['search_time_to_reward'][-1][2] print 'Time: %.2f ' % tottime sys.exit(-1) set_seed(parameters.pidx) problem_env = PaPMoverEnv(parameters.pidx) goal_objs = [ 'square_packing_box1', 'square_packing_box2', 'rectangular_packing_box3', 'rectangular_packing_box4' ] goal_region = 'home_region' problem_env.set_goal(goal_objs, goal_region) goal_entities = goal_objs + [goal_region] if parameters.use_shaped_reward: # uses the reward shaping per Ng et al. reward_function = ShapedRewardFunction(problem_env, goal_objs, goal_region, parameters.planning_horizon) else: reward_function = GenericRewardFunction(problem_env, goal_objs, goal_region, parameters.planning_horizon) motion_planner = OperatorBaseMotionPlanner(problem_env, 'prm') problem_env.set_reward_function(reward_function) problem_env.set_motion_planner(motion_planner) learned_q = None prior_q = None if parameters.use_learned_q: learned_q = load_learned_q(parameters, problem_env) v_fcn = lambda state: -len(get_objects_to_move(state, problem_env)) if parameters.planner == 'mcts': planner = MCTS(parameters, problem_env, goal_entities, prior_q, learned_q) elif parameters.planner == 'mcts_with_leaf_strategy': planner = MCTSWithLeafStrategy(parameters, problem_env, goal_entities, v_fcn, learned_q) else: raise NotImplementedError set_seed(parameters.planner_seed) stime = time.time() search_time_to_reward, n_feasibility_checks, plan = planner.search( max_time=parameters.timelimit) tottime = time.time() - stime print 'Time: %.2f ' % (tottime) # todo # save the entire tree pickle.dump( { "search_time_to_reward": search_time_to_reward, 'plan': plan, 'commit_hash': commit_hash, 'n_feasibility_checks': n_feasibility_checks, 'n_nodes': len(planner.tree.get_discrete_nodes()) }, open(save_dir + filename, 'wb'))