def precompute_and_save_intersections(t, enf_locs, trace): print "Precomputing intersections..." intersection_cache = [] for i in xrange(t+1): next_enf_loc = scale_up(enf_locs[i]) cur_enf_loc = scale_up(enf_locs[i-1]) fv = direction(next_enf_loc, cur_enf_loc) intersections = trace.model.isovist.GetIsovistIntersections(cur_enf_loc, fv) intersection_cache.append(intersections) cPickle.dump( intersection_cache, open("./enf_int_cache.cp","w") ) print "done!, Interction Count:", len(intersection_cache) return intersection_cache
def run_smart(self, Q): t = Q.choice( p=1.0/40*np.ones((1,40)), name="t" ) # current location of chaser (at time 't') curr_loc = [Q.get_obs("enf_x_"+str(t)), Q.get_obs("enf_y_"+str(t))] # randomly choose a start and goal for runner cnt = len(self.locs) int_start_i = Q.choice( p=1.0/cnt*np.ones((1,cnt)), name="int_start" ) int_goal_i = Q.choice( p=1.0/cnt*np.ones((1,cnt)), name="int_goal" ) # create empty trace using model q = p.ProgramTrace(self.runner_model) # condition q on Q q = self.condition_q(Q, q) # run inference to get intruder's expected next step post_sample_traces = run_inference(q, post_samples=6, samples=5) # 10, 5 runner_exp_next_step, runner_exp_path = rand_expected_future_step(post_sample_traces, "int_plan") #if point_in_obstacle(runner_exp_next_step, self.epolys): # runner_exp_next_step = get_clear_goal(curr_loc, runner_exp_next_step, self.polys) Q.keep("q", q.trace) Q.keep("runner_post_sample_traces", post_sample_traces) Q.keep("runner_exp_next_step", runner_exp_next_step) # plan path toward anticipated location (t+1) for runner #enf_plan = self.plan_path(np.atleast_2d( curr_loc), np.atleast_2d( runner_exp_next_step)) rx1,ry1,rx2,ry2 = self.seg_map enf_plan = planner.run_rrt_opt( np.atleast_2d( curr_loc), np.atleast_2d( runner_exp_next_step), rx1,ry1,rx2,ry2) Q.keep("enf_plan", enf_plan) # set up the enforcer view (forward vector, fv) for the next step cur_enf_loc = scale_up(curr_loc) next_enf_loc = scale_up(enf_plan[1]) next_int_loc = scale_up(runner_exp_path[t+1]) fv = direction(next_int_loc, cur_enf_loc) intersections = self.isovist.GetIsovistIntersections(next_enf_loc, fv) # does the enforcer see me at time 't' #runner_next_loc = scale_up(runner_exp_next_step) will_runner_be_seen = self.isovist.FindIntruderAtPoint( next_int_loc, intersections ) detected_prob = 0.999*will_runner_be_seen + 0.001*(1-will_runner_be_seen) # ~ flip(seen*.999 + (1-seen*.001) # XXX should consider seeing if the enforcer will see the intruder # before reaching its simulated goal location, not just in the next step # down side: it would need more isovist calculations at each step runner_detected = Q.flip( p=detected_prob, name="int_detected" )
def run_naive(self, Q): t = Q.choice( p=1.0/40*np.ones((1,40)), name="t" ) # current location of chaser (at time 't') curr_loc = [Q.get_obs("enf_x_"+str(t)), Q.get_obs("enf_y_"+str(t))] # randomly choose a start and goal for runner cnt = len(self.locs) int_start_i = Q.choice( p=1.0/cnt*np.ones((1,cnt)), name="int_start" ) int_goal_i = Q.choice( p=1.0/cnt*np.ones((1,cnt)), name="int_goal" ) rx1,ry1,rx2,ry2 = self.seg_map int_plan = planner.run_rrt_opt( np.atleast_2d(self.locs[int_start_i]), np.atleast_2d( self.locs[int_goal_i]), rx1,ry1,rx2,ry2, slow=True) # plan path toward anticipated location (t+1) for runner #enf_plan = self.plan_path(np.atleast_2d( curr_loc), np.atleast_2d( runner_exp_next_step)) t_ = min(t+12, len(int_plan)-1) enf_plan = planner.run_rrt_opt( np.atleast_2d( curr_loc), np.atleast_2d( int_plan[t_]), rx1,ry1,rx2,ry2) Q.keep("enf_plan", enf_plan) Q.keep("int_plan", int_plan) # set up the enforcer view (forward vector, fv) for the next step cur_enf_loc = scale_up(curr_loc) next_enf_loc = scale_up(enf_plan[1]) fv = direction(next_enf_loc, cur_enf_loc) intersections = self.isovist.GetIsovistIntersections(next_enf_loc, fv) # does the enforcer see me at time 't' runner_next_loc = scale_up(int_plan[t+1]) will_runner_be_seen = self.isovist.FindIntruderAtPoint( int_plan[t+1], intersections ) detected_prob = 0.999*will_runner_be_seen + 0.001*(1-will_runner_be_seen) # ~ flip(seen*.999 + (1-seen*.001) # XXX should consider seeing if the enforcer will see the intruder # before reaching its simulated goal location, not just in the next step # down side: it would need more isovist calculations at each step runner_detected = Q.flip( p=detected_prob, name="int_detected" )
def run(self, Q): #---------------------------------------------------------- # simplified enforcer model #---------------------------------------------------------- t = Q.choice(p=1.0 / 40 * np.ones((1, 40)), name="t") cnt = len(self.locs) enf_start_i = Q.choice(p=1.0 / cnt * np.ones((1, cnt)), name="enf_start") enf_goal_i = Q.choice(p=1.0 / cnt * np.ones((1, cnt)), name="enf_goal") enf_start = np.atleast_2d(self.locs[enf_start_i]) enf_goal = np.atleast_2d(self.locs[enf_goal_i]) path_noise = .003 #enf_plan = self.plan_path(enf_start, enf_goal) rx1, ry1, rx2, ry2 = self.seg_map enf_plan = planner.run_rrt_opt(np.atleast_2d(enf_start), np.atleast_2d(enf_goal), rx1, ry1, rx2, ry2) enf_noisy_plan = [enf_plan[0]] for i in xrange( 1, len(enf_plan) - 1 ): #loc_t = np.random.multivariate_normal(enf_plan[i], [[path_noise, 0], [0, path_noise]]) # name 't_i' i.e. t_1, t_2,...t_n loc_x = Q.randn(mu=enf_plan[i][0], sigma=path_noise, name="enf_x_" + str(i)) loc_y = Q.randn(mu=enf_plan[i][1], sigma=path_noise, name="enf_y_" + str(i)) loc_t = [loc_x, loc_y] enf_noisy_plan.append(loc_t) enf_noisy_plan.append(enf_plan[-1]) enf_loc = np.atleast_2d(enf_noisy_plan[t]) #----------------- end of enforcer model ------------------ # XXX the runner wants to do goal inference to figure out the next step of the enf # TWO desires: # 1) high likelihood for a enf path conditioned on past locations # 2) high likelihood for his own path conditioned of not being detected for the next step and conditioned # on past enf locations ! #---------------------------------------------------------- # runner (intruder) model #---------------------------------------------------------- start_i = Q.choice(p=1.0 / cnt * np.ones((1, cnt)), name="int_start") goal_i = Q.choice(p=1.0 / cnt * np.ones((1, cnt)), name="int_goal") start = np.atleast_2d(self.locs[start_i]) goal = np.atleast_2d(self.locs[goal_i]) #my_plan = self.plan_path(start, goal) my_plan = planner.run_rrt_opt(np.atleast_2d(start), np.atleast_2d(goal), rx1, ry1, rx2, ry2) my_loc = np.atleast_2d(my_plan[t]) my_noisy_plan = [my_plan[0]] for i in xrange( 1, len(my_plan) - 1 ): #loc_t = np.random.multivariate_normal(my_plan[i], [[path_noise, 0], [0, path_noise]]) # name 't_i' i.e. t_1, t_2,...t_n loc_x = Q.randn(mu=my_plan[i][0], sigma=path_noise, name="int_x_" + str(i)) loc_y = Q.randn(mu=my_plan[i][1], sigma=path_noise, name="int_y_" + str(i)) loc_t = [loc_x, loc_y] my_noisy_plan.append(loc_t) my_noisy_plan.append(my_plan[-1]) # XXX Need to make sure the runner wasn't seen in any of the previous time steps i_already_seen = 0 for i in xrange(t): intersections = Q.cache["enf_intersections_t_" + str(i)] # get enforcer's fv for time t i_already_seen = self.isovist.FindIntruderAtPoint( my_noisy_plan[i], intersections) if (i_already_seen): detected_prob = 0.999 * i_already_seen + 0.001 * ( 1 - i_already_seen) if not i_already_seen: # set up the enforcer view (forward vector, fv) for the next step cur_enf_loc = scale_up(enf_noisy_plan[t]) next_enf_loc = scale_up(enf_noisy_plan[t + 1]) fv = direction(next_enf_loc, cur_enf_loc) intersections = self.isovist.GetIsovistIntersections( next_enf_loc, fv) # does the enforcer see me at time 't' my_next_loc = scale_up(my_noisy_plan[t + 1]) will_i_be_seen = self.isovist.FindIntruderAtPoint( my_next_loc, intersections) detected_prob = 0.999 * will_i_be_seen + 0.001 * ( 1 - will_i_be_seen) # ~ flip(seen*.999 + (1-seen*.001) future_detection = Q.flip(p=detected_prob, name="int_detected") # for rendering purposes Q.keep("int_plan", my_noisy_plan) Q.keep("enf_plan", enf_noisy_plan)
def run_basic_partial(self, Q, path_noise=0.003): rx1, ry1, rx2, ry2 = self.seg_map t = Q.choice(p=1.0 / 40 * np.ones((1, 40)), name="t") # #------------- model other agent movements (past and future) -------------- XXX need to change RV names o_start_i = Q.choice(p=1.0 / self.cnt * np.ones((1, self.cnt)), name="other_run_start") o_goal_i = Q.choice(p=1.0 / self.cnt * np.ones((1, self.cnt)), name="other_run_goal") o_start = np.atleast_2d(self.locs[o_start_i]) o_goal = np.atleast_2d(self.locs[o_goal_i]) # plan using the latent variables of start and goal other_plan = planner.run_rrt_opt(np.atleast_2d(o_start), np.atleast_2d(o_goal), rx1, ry1, rx2, ry2) # add noise to the plan other_noisy_plan = [other_plan[0]] for i in xrange( 1, len(other_plan) - 1 ): #loc_t = np.random.multivariate_normal(my_plan[i], [[path_noise, 0], [0, path_noise]]) # name 't_i' i.e. t_1, t_2,...t_n loc_x = Q.randn(mu=other_plan[i][0], sigma=path_noise, name="other_run_x_" + str(i)) loc_y = Q.randn(mu=other_plan[i][1], sigma=path_noise, name="other_run_y_" + str(i)) loc_t = [loc_x, loc_y] other_noisy_plan.append(loc_t) other_noisy_plan.append(other_plan[-1]) #------------- model agent's own movements (past and future) -------------- start_i = Q.choice(p=1.0 / self.cnt * np.ones((1, self.cnt)), name="run_start") goal_i = Q.choice(p=1.0 / self.cnt * np.ones((1, self.cnt)), name="run_goal") start = np.atleast_2d(self.locs[start_i]) goal = np.atleast_2d(self.locs[goal_i]) # plan using the latent variables of start and goal my_plan = planner.run_rrt_opt(np.atleast_2d(start), np.atleast_2d(goal), rx1, ry1, rx2, ry2) # add noise to the plan my_noisy_plan = [my_plan[0]] for i in xrange( 1, len(my_plan) - 1 ): #loc_t = np.random.multivariate_normal(my_plan[i], [[path_noise, 0], [0, path_noise]]) # name 't_i' i.e. t_1, t_2,...t_n loc_x = Q.randn(mu=my_plan[i][0], sigma=path_noise, name="run_x_" + str(i)) loc_y = Q.randn(mu=my_plan[i][1], sigma=path_noise, name="run_y_" + str(i)) loc_t = [loc_x, loc_y] my_noisy_plan.append(loc_t) my_noisy_plan.append(my_plan[-1]) my_loc = my_noisy_plan[t] #---------------- need to add RV of detection for each time step ---------- t_detected = [] PATH_LIMIT = 40 for i in xrange(0, PATH_LIMIT): cur_loc = scale_up(my_noisy_plan[i]) intersections = None detection_prob = 0.001 # face the runner if within certain radius if dist(my_noisy_plan[i], other_noisy_plan[i]) <= .4: #.35: fv = direction(scale_up(other_noisy_plan[i]), cur_loc) intersections = self.isovist.GetIsovistIntersections( cur_loc, fv) # does the agent see other at time 't' other_loc = scale_up(other_noisy_plan[i]) will_other_be_seen = self.isovist.FindIntruderAtPoint( other_loc, intersections) if will_other_be_seen: detection_prob = 0.999 t_detected.append(i) future_detection = Q.flip(p=detection_prob, name="detected_t_" + str(i)) Q.keep("intersections-t-" + str(i), intersections) same_goal = 0 if goal_i == o_goal_i: same_goal = 1 same_goal_prob = 0.999 * same_goal + 0.001 * (1 - same_goal) runners_same_goal = Q.flip(p=same_goal_prob, name="same_goal") Q.keep("t_detected", t_detected) Q.keep("my_plan", my_noisy_plan) Q.keep("other_plan", other_noisy_plan)
def run_simulation(sim_id, locs, seg_map, isovist, polys, epolys): # unpack rx1, ry1, rx2, ry2 = seg_map intersection_cache = [] # create model of chaser chaser_model = create_chaser_model(seg_map=seg_map, locs=locs, isovist=isovist) # get evenly spead out starts and goal int_start = 0 int_goal = 8 #chas_start = 1 # int_start = randint(0, len(locs)-1) # int_goal = randint(0, len(locs)-1) # while(dist(locs[int_start], locs[int_goal]) <= 0.5): # int_goal = randint(0, len(locs)-1) chas_start = randint(0, len(locs) - 1) while (dist(locs[int_start], locs[chas_start]) <= 0.5): chas_start = randint(0, len(locs) - 1) runner_start_i = int_start runner_start = locs[runner_start_i] # get runner's random goal location (far enough from its start) runner_goal_i = int_goal runner_goal = locs[runner_goal_i] # store history of runner's locations runner_true_locs = [runner_start] # get random start location start_index = chas_start chaser_start = locs[start_index] # begin tracking history of movements chaser_locs = [chaser_start] #TEMP FOR TESTING (NAIVE RUNNER) runner_path = planner.run_rrt_opt(np.atleast_2d(runner_start), np.atleast_2d(runner_goal), rx1, ry1, rx2, ry2, slow=True) if not naive_runner: # XXX smart runner runner_path = [[0.10000000000000001, 0.099999999999999978], [0.14370786380607686, 0.12428214655893158], [0.18741572761215378, 0.14856429311786318], [0.23112359141823066, 0.17284643967679481], [0.27483145522430752, 0.19712858623572641], [0.25854686086588807, 0.23830235506087813], [0.25097278902994408, 0.28619750551848688], [0.24540487293466112, 0.31648310191196649], [0.19679616003865907, 0.30477015904546001], [0.16866150321818352, 0.32844002322014876], [0.15786536899041551, 0.37717509770481417], [0.14948852463283413, 0.44046838810633692], [0.14111168027525278, 0.47576167850785966], [0.13273483591767141, 0.52505496890938252], [0.12435799156009002, 0.56434825931090527], [0.11598114720250866, 0.62364154971242802], [0.13916939803284978, 0.66269402652378151], [0.15039748869499225, 0.70969377654572163], [0.15619668605039488, 0.75935633095075694], [0.16199588340579754, 0.80901888535579236], [0.16779508076120014, 0.85868143976082756], [0.20837094819244148, 0.87335725908849149], [0.25828353694770523, 0.87588922948695491], [0.30825277792753181, 0.87764278470422896], [0.35822201890735811, 0.87939633992150279], [0.40819125988718469, 0.88114989513877684], [0.45816050086701121, 0.88290345035605078], [0.48130593126207499, 0.91889699806900549], [0.53128113014992562, 0.92047163965355017], [0.58125632903777602, 0.92204628123809473]] for i in xrange(10): runner_path.append(runner_path[len(runner_path) - 1]) runner_goal = runner_path[-1] #--end of smart runner assert len(runner_path) == 40 runner_locs = runner_path Q_history = [] Q_history_ts = [] # begin timer TIME_LIMIT = 35 for t in xrange(TIME_LIMIT): # check if the runner has reached its goal try: if runner_locs[t + 1] == runner_goal: print "Failed: Runner Reached Goal" return False, Q_history_ts break except: print "FAILED" return False, Q_history_ts # get true runner's location at time t + 1 runner_true_loc = runner_locs[t + 1] runner_true_locs.append(runner_true_loc) #--------------------------------- # create empty trace Q = p.ProgramTrace(chaser_model) # condition the trace Q.condition("t", t) Q.set_obs("enf_start", chas_start) Q.condition("int_detected", True) Q.condition("int_start", runner_start_i) for pre_t in xrange(t + 1): Q.set_obs("enf_x_" + str(pre_t), chaser_locs[pre_t][0]) Q.set_obs("enf_y_" + str(pre_t), chaser_locs[pre_t][1]) if pre_t < t: Q.cache["enf_intersections_t_" + str(pre_t)] = intersection_cache[pre_t] # run inference if naive_chaser: post_sample_traces = run_inference(Q, post_samples=64, samples=32) else: post_sample_traces = run_inference(Q, post_samples=64, samples=32) # nested inference exp_next_step, exp_enf_path = rand_expected_future_step( post_sample_traces, "enf_plan") #XXX not working #if point_in_obstacle(chaser_locs[t], epolys): # print "chaser went inside building" #return False Q.keep("post_sample_traces", post_sample_traces) # replan to its expected next step #t_ = min(t, len()) # enf_plan = planner.run_rrt_opt( np.atleast_2d(chaser_locs[t]), # np.atleast_2d(exp_next_step), rx1,ry1,rx2,ry2) enf_plan = exp_enf_path # ------------------------------------ if enf_plan is None: print "trying again" t -= 1 continue enf_next_step = enf_plan[1] # store step made chaser_locs.append(enf_next_step) inters = None # check to see if the runner was detected runner_detected = False # only if he is close if dist(enf_next_step, runner_true_loc) <= .4: # add a bit of hearing by facing the chaser towards the runner fv = direction(scale_up(runner_true_loc), scale_up(enf_next_step)) # check if runner can be detected runner_true_next_loc = scale_up(runner_true_loc) intersections = isovist.GetIsovistIntersections( scale_up(enf_next_step), fv) runner_detected = isovist.FindIntruderAtPoint( scale_up(runner_true_loc), intersections) inters = np.asarray(intersections) # store intersections at time 't' facing in the direction it stepped fv = direction(scale_up(chaser_locs[t + 1]), scale_up(chaser_locs[t])) intersections = isovist.GetIsovistIntersections( scale_up(chaser_locs[t + 1]), fv) if not runner_detected: runner_detected = isovist.FindIntruderAtPoint( scale_up(runner_true_loc), intersections) intersection_cache.append(intersections) if inters is None: inters = np.asarray(intersections) inters /= 500.0 # plot behavior plot_name = None if runner_detected: print "Success: Runner Detected Before Goal" plot_name = str(sim_id) + "_t-" + str(t) + "_s-" + "T" + ".eps" if naive_chaser: save_chaser_post_traces_naive( post_sample_traces, plot_name=plot_name, runner_true_locs=runner_true_locs, chaser_true_locs=chaser_locs, intersections=inters, exp_enf_path=exp_enf_path) else: save_chaser_post_traces_nested( post_sample_traces, plot_name=plot_name, runner_true_locs=runner_true_locs, chaser_true_locs=chaser_locs, intersections=inters, exp_enf_path=exp_enf_path) Q.keep("real_world_detection", True) Q_history.append(Q) Q_history_ts.append(Q.trace) return True, Q_history_ts else: plot_name = str(sim_id) + "_t-" + str(t) + "_s-" + "F" + ".eps" Q.keep("real_world_detection", False) Q_history.append(Q) Q_history_ts.append(Q.trace) print "searching..." if naive_chaser: save_chaser_post_traces_naive(post_sample_traces, plot_name=plot_name, runner_true_locs=runner_true_locs, chaser_true_locs=chaser_locs, intersections=inters, exp_enf_path=exp_enf_path) else: save_chaser_post_traces_nested(post_sample_traces, plot_name=plot_name, runner_true_locs=runner_true_locs, chaser_true_locs=chaser_locs, intersections=inters, exp_enf_path=exp_enf_path) print "Failed: Runner Reached Goal" return False, Q_history_ts
def run_tom_partial(self, Q, path_noise=0.003): rx1,ry1,rx2,ry2 = self.seg_map t = Q.choice( p=1.0/40*np.ones((1,40)), name="t" ) #------------- model agent's movements (past and future) -------------- start_i = Q.choice( p=1.0/self.cnt*np.ones((1,self.cnt)), name="init_run_start" ) goal_i = Q.choice( p=1.0/self.cnt*np.ones((1,self.cnt)), name="init_run_goal" ) start = np.atleast_2d( self.locs[start_i] ) goal = np.atleast_2d( self.locs[goal_i] ) # plan using the latent variables of start and goal my_plan = planner.run_rrt_opt( np.atleast_2d(start), np.atleast_2d(goal), rx1,ry1,rx2,ry2 ) # add noise to the plan my_noisy_plan = [my_plan[0]] for i in xrange(1, len(my_plan)-1):#loc_t = np.random.multivariate_normal(my_plan[i], [[path_noise, 0], [0, path_noise]]) # name 't_i' i.e. t_1, t_2,...t_n loc_x = Q.randn( mu=my_plan[i][0], sigma=path_noise, name="init_run_x_"+str(i) ) loc_y = Q.randn( mu=my_plan[i][1], sigma=path_noise, name="init_run_y_"+str(i) ) loc_t = [loc_x, loc_y] my_noisy_plan.append(loc_t) my_noisy_plan.append(my_plan[-1]) my_loc = my_noisy_plan[t] #---------------- do inference -------------------------------------------- other_noisy_plan = None other_inferred_trace = None if self.mode == "collab": post_sample_traces, other_inferred_goal = self.collaborative_nested_inference(Q) for trace in post_sample_traces: if trace["run_goal"] == other_inferred_goal: other_inferred_trace = trace other_noisy_plan = trace["my_plan"] break if self.mode == "advers": # return the trace with least detections nested_post_sample_traces, other_inferred_trace = self.adversarial_nested_inference(Q) other_noisy_plan = other_inferred_trace["my_plan"] #---------------- need to add RV of detection for each time step ---------- t_detected = [] PATH_LIMIT = 40 for i in xrange(0, PATH_LIMIT): cur_loc = scale_up(my_noisy_plan[i]) intersections = None detection_prob = -10000.0 # face the runner if within certain radius if dist(my_noisy_plan[i], other_noisy_plan[i]) <= .4: #.35: fv = direction(scale_up(other_noisy_plan[i]), cur_loc) intersections = self.isovist.GetIsovistIntersections(cur_loc, fv) # does the agent see other at time 't' other_loc = scale_up(other_noisy_plan[i]) will_other_be_seen = self.isovist.FindIntruderAtPoint( other_loc, intersections ) if will_other_be_seen: detection_prob = -.01 t_detected.append(i) future_detection = Q.lflip( lp=detection_prob, name="detected_t_"+str(i) ) Q.keep("intersections-t-"+str(i), intersections) if self.mode == "collab": same_goal = 0 if goal_i == other_inferred_goal: same_goal = 1 same_goal_prob = 0.999*same_goal + 0.001*(1-same_goal) runners_same_goal = Q.flip( p=same_goal_prob, name="same_goal" ) #TODO: update this lflip 1.19.2018 #print t_detected Q.keep("t_detected", t_detected) Q.keep("my_plan", my_noisy_plan) Q.keep("other_plan", other_noisy_plan) Q.keep("other_run_start", other_inferred_trace["run_start"]) Q.keep("other_run_goal", other_inferred_trace["run_goal"]) Q.keep("nested_post_samples", nested_post_sample_traces)
def run_basic_partial(self, Q, path_noise=0.000): rx1,ry1,rx2,ry2 = self.seg_map t = Q.choice( p=1.0/40*np.ones((1,40)), name="t" ) # #------------- model other agent movements (past and future) -------------- o_start_i = Q.choice( p=1.0/self.cnt*np.ones((1,self.cnt)), name="other_run_start" ) o_goal_i = Q.choice( p=1.0/self.cnt*np.ones((1,self.cnt)), name="other_run_goal" ) o_start = np.atleast_2d( self.locs[o_start_i] ) o_goal = np.atleast_2d( self.locs[o_goal_i] ) # plan using the latent variables of start and goal # other_plan = planner.run_rrt_opt( np.atleast_2d(o_start), # np.atleast_2d(o_goal), rx1,ry1,rx2,ry2 ) #print Q.fetch_obs() prev_true_loc = np.atleast_2d([Q.get_obs("other_run_x_"+str(t-1)), Q.get_obs("other_run_y_"+str(t-1))]) #print "prev_true_loc", prev_true_loc other_plan = planner.run_rrt_opt( prev_true_loc, np.atleast_2d(o_goal), rx1,ry1,rx2,ry2 ) u = 1 # add noise to the plan other_noisy_plan = [self.locs[o_start_i]] for i in xrange(1, len(other_plan)-1):#loc_t = np.random.multivariate_normal(my_plan[i], [[path_noise, 0], [0, path_noise]]) # name 't_i' i.e. t_1, t_2,...t_n if i < t: # do not sample if already known to be true: loc_t = [Q.get_obs(name="other_run_x_"+str(i)), Q.get_obs(name="other_run_y_"+str(i))] else: loc_x = Q.randn( mu=other_plan[u][0], sigma=path_noise, name="other_run_x_"+str(i) ) loc_y = Q.randn( mu=other_plan[u][1], sigma=path_noise, name="other_run_y_"+str(i) ) loc_t = [loc_x, loc_y] u+=1 other_noisy_plan.append(loc_t) other_noisy_plan.append(other_plan[-1]) #------------- model agent's own movements (past and future) -------------- start_i = Q.choice( p=1.0/self.cnt*np.ones((1,self.cnt)), name="run_start" ) goal_i = Q.choice( p=1.0/self.cnt*np.ones((1,self.cnt)), name="run_goal" ) #insert as a way to get goals not too close goal_i, Q = get_goal(start_i, Q) start = np.atleast_2d( self.locs[start_i] ) goal = np.atleast_2d( self.locs[goal_i] ) # plan using the latent variables of start and goal my_plan = planner.run_rrt_opt( np.atleast_2d(start), np.atleast_2d(goal), rx1,ry1,rx2,ry2 ) # add noise to the plan my_noisy_plan = [my_plan[0]] for i in xrange(1, len(my_plan)-1):#loc_t = np.random.multivariate_normal(my_plan[i], [[path_noise, 0], [0, path_noise]]) # name 't_i' i.e. t_1, t_2,...t_n loc_x = Q.randn( mu=my_plan[i][0], sigma=path_noise, name="run_x_"+str(i) ) loc_y = Q.randn( mu=my_plan[i][1], sigma=path_noise, name="run_y_"+str(i) ) loc_t = [loc_x, loc_y] my_noisy_plan.append(loc_t) my_noisy_plan.append(my_plan[-1]) my_loc = my_noisy_plan[t] #-------------------------- detections --------------------- if self.mode == "collab": t_detected = [] PATH_LIMIT = 30 for i in xrange(0, PATH_LIMIT): cur_loc = scale_up(my_noisy_plan[i]) intersections = None detection_prob = 0.1 # face the runner if within certain radius if dist(my_noisy_plan[i], other_noisy_plan[i]) <= .4: #.35: fv = direction(scale_up(other_noisy_plan[i]), cur_loc) intersections = self.isovist.GetIsovistIntersections(cur_loc, fv) # does the agent see other at time 't' other_loc = scale_up(other_noisy_plan[i]) will_other_be_seen = self.isovist.FindIntruderAtPoint( other_loc, intersections ) if will_other_be_seen: detection_prob = 0.9 t_detected.append(i) future_detection = Q.flip( p=detection_prob, name="detected_t_"+str(i) ) # ^ might want to change this to lflip when you do experiments next # you wrote this on 1.19.2018 - iris Q.keep("intersections-t-"+str(i), intersections) same_goal = 0 if goal_i == o_goal_i: same_goal = 1 same_goal_prob = .999*same_goal + .001*(1-same_goal) runners_same_goal = Q.lflip( lp=same_goal_prob, name="same_goal" ) Q.keep("t_detected", t_detected) Q.keep("my_plan", my_noisy_plan) Q.keep("other_plan", other_noisy_plan) if self.mode == "advers": t_detected = [] PATH_LIMIT = 30 for i in xrange(0, PATH_LIMIT): other_loc = scale_up(other_noisy_plan[i]) intersections = None detection_prob = -0.0001 # face the runner if within certain radius if dist(my_noisy_plan[i], other_noisy_plan[i]) <= .4: #.35: # -----if other is looking for me #print ("fv:", scale_up(my_noisy_plan[i]), other_loc) fv = direction(scale_up(my_noisy_plan[i]), other_loc) intersections = self.isovist.GetIsovistIntersections(other_loc, fv) #print intersections # does the chaser see other at time 'i' curr_loc = scale_up(my_noisy_plan[i]) #if other is looking for me will_I_be_seen = self.isovist.FindIntruderAtPoint( curr_loc, intersections ) #print will_I_be_seen if will_I_be_seen: detection_prob = -1.0 t_detected.append(i) future_detection = Q.clflip( lp=detection_prob, name="detected_t_"+str(i) ) #future_detection = Q.flip( p=detection_prob, name="detected_t_"+str(i) ) Q.keep("intersections-t-"+str(i), intersections) Q.keep("t_detected", t_detected) Q.keep("my_plan", my_noisy_plan) Q.keep("other_plan", other_noisy_plan)
def run_tom_partial(self, Q, path_noise=0.000): rx1,ry1,rx2,ry2 = self.seg_map t = Q.choice( p=1.0/40*np.ones((1,40)), name="t" ) #------------- model agent's movements (past and future) -------------- start_i = Q.choice( p=1.0/self.cnt*np.ones((1,self.cnt)), name="init_run_start" ) goal_i = Q.choice( p=1.0/self.cnt*np.ones((1,self.cnt)), name="init_run_goal" ) start = np.atleast_2d( self.locs[start_i] ) goal = np.atleast_2d( self.locs[goal_i] ) #print "HERE:", goal_i, goal # plan using the latent variables of start and goal prev_true_loc = np.atleast_2d([Q.get_obs("init_run_x_"+str(t-1)), Q.get_obs("init_run_y_"+str(t-1))]) my_plan = planner.run_rrt_opt( prev_true_loc, goal, rx1,ry1,rx2,ry2 ) u = 1 my_noisy_plan = [self.locs[start_i]] for i in xrange(1, len(my_plan)-1):#loc_t = np.random.multivariate_normal(my_plan[i], [[path_noise, 0], [0, path_noise]]) # name 't_i' i.e. t_1, t_2,...t_n if i < t: # do not sample if already known to be true loc_t = [Q.get_obs("init_run_x_"+str(i)), Q.get_obs("init_run_y_"+str(i))] else: loc_x = Q.randn( mu=my_plan[u][0], sigma=path_noise, name="init_run_x_"+str(i) ) loc_y = Q.randn( mu=my_plan[u][1], sigma=path_noise, name="init_run_y_"+str(i) ) loc_t = [loc_x, loc_y] u += 1 my_noisy_plan.append(loc_t) my_noisy_plan.append(my_plan[-1]) my_loc = my_noisy_plan[t] #assert (len(my_noisy_plan) == 30) #---------------- do inference -------------------------------------------- other_noisy_plan = None other_inferred_trace = None if self.mode == "collab": post_sample_traces, other_inferred_goal = self.collaborative_nested_inference(Q) for trace in post_sample_traces: if trace["run_goal"] == other_inferred_goal: other_inferred_trace = trace other_noisy_plan = trace["my_plan"] break if self.mode == "advers": all_t_detected = [] other_plans = [] L = self.L all_Qls_scores = [] all_Qls_traces = [] all_Qls_obs = [] all_ql_scores = [] #print "init Q score:", Q.get_score() for l in xrange(L): q = self.condition_middle_model(Q) # print ("Q obs:", Q.fetch_obs()) # print ("q obs:", q.fetch_obs()) # raw_input() Q_l = Q.get_copy() #print "---------running runner model--------------" q_score_l, q_l = q.run_model() #print '----------running rest of chaser model-----------' #plot_middlemost_sample(q_l, q_score_l, self.directory) other_noisy_plan = q_l["my_plan"] #---------------- need to add RV of detection for each time step ---------- t_detected = [] PATH_LIMIT = 30 will_other_be_seen = False for i in xrange(0, PATH_LIMIT): cur_loc = scale_up(my_noisy_plan[i]) intersections = None detection_prob = -.0001 # face the runner if within certain radius if dist(my_noisy_plan[i], other_noisy_plan[i]) <= .4: #.35: fv = direction(scale_up(other_noisy_plan[i]), cur_loc) intersections = self.isovist.GetIsovistIntersections(cur_loc, fv) # does the agent see other at time 't' other_loc = scale_up(other_noisy_plan[i]) will_other_be_seen = self.isovist.FindIntruderAtPoint( other_loc, intersections ) if will_other_be_seen: detection_prob = -1 t_detected.append(i) #future_detection = Q_l.flip( p=detection_prob, name="detected_t_"+str(i) ) future_detection = Q_l.lflip( lp=detection_prob, name="detected_t_"+str(i)) Q_l.add_trace(name="q_trace", trace=q_l, score=q_score_l) all_ql_scores.append(q_score_l) #print "list:", all_ql_scores # Q_l.keep("t_detected", t_detected) all_t_detected.append(t_detected) # Q_l.keep("my_plan", my_noisy_plan) # Q_l.keep("other_plan", other_noisy_plan) other_plans.append(other_noisy_plan) # Q_l.keep("other_run_start", q_l["run_start"]) # Q_l.keep("other_run_goal", q_l["run_goal"]) all_Qls_scores.append(Q_l.get_score()) all_Qls_traces.append(Q_l.get_trace()) all_Qls_obs.append(Q_l.get_obs()) #print ("all ql scores:", all_ql_scores) #print ("all_Qls_scores:", all_Qls_scores) Q.keep("all_ql_scores", np.array(all_ql_scores)) Q.keep("all_Qls_obs", all_Qls_obs) Q.keep("all_Qls_traces", all_Qls_traces) Q.keep("all_Qls_scores", np.array(all_Qls_scores)) Q.keep("mean", np.mean(all_Qls_scores)) Q.keep("my_plan", my_noisy_plan) Q.keep("t_detected", all_t_detected) Q.keep("other_plan", other_plans) #update actual score of Q. Q.cur_trace_score = np.mean(all_Qls_scores)