def static_control(track_select='SS'):
Environment.track_generator(track, track_select=track_select)
env = Environment.Environment(track, 10000)
gui = GUI.GUI(track, cars, trace=True)
car_objects = [Environment.Car(c) for c in cars]
env.compute_interaction(car_objects)
GOALS = ((0.33, 0), (0.14, 0.6), (0.4, 0), (0.148, -0.6))
while (True):
for goal in GOALS:
for i in range(len(car_objects)):
car_objects[i].set_velocity(goal[0])
car_objects[i].set_steering(goal[1])
for i in range(100):
debug_data = ''
for i in range(len(car_objects)):
if car_objects[i].state == 'collided':
debug_data += 'Car ' + str(i) + '\nCollided!\n\n'
continue
car_objects[i].update(dt)
s_r = car_objects[i].get_sensor_reading()
gui.update(i, car_objects[i].get_state())
debug_data += 'Car ' + str(
i) + '\nSensor readings:' + ', '.join(
['{:.2f}'.format(x)
for x in s_r]) + '\nCar score=' + '{:.2f}'.format(
car_objects[i].score) + '\n'
env.compute_interaction(car_objects)
gui.update_debug_info(debug_data)
gui.refresh()
def user_control(track_select='SS'):
Environment.track_generator(track, track_select=track_select)
env = Environment.Environment(track, 10000)
gui = GUI.GUI(track, cars, trace=True)
car_objects = [Environment.Car(c) for c in cars]
env.compute_interaction(car_objects)
while (True):
d = {
'w': [0.5, 0.0],
's': [-0.3, 0.0],
'a': [0.1, 0.6],
'd': [0.1, -0.6]
}
try:
[v, s] = d[raw_input()]
except:
[v, s] = [0, 0]
for agent in car_objects:
agent.set_velocity(v)
agent.set_steering(s)
for i in range(10):
debug_data = ''
for i in range(len(car_objects)):
if car_objects[i].state == 'collided':
debug_data += 'Car ' + str(i) + '\nCollided!\n\n'
continue
car_objects[i].update(dt)
s_r = car_objects[i].get_sensor_reading()
gui.update(i, car_objects[i].get_state())
debug_data += 'Car ' + str(
i) + '\nSensor readings:' + ', '.join(
['{:.2f}'.format(x)
for x in s_r]) + '\nCar score=' + '{:.2f}'.format(
car_objects[i].score) + '\n'
env.compute_interaction(car_objects)
gui.update_debug_info(debug_data)
gui.refresh()
def reinfrocement_neural_network_control(load_weights=None,
run_only=False,
track_select='SS',
random_seed=None,
rl_prams=None):
run = run_only
weights_save_dir = "./weights/"
if not os.path.exists(weights_save_dir): os.makedirs(weights_save_dir)
Environment.track_generator(track, track_select=track_select)
env = Environment.Environment(track, rl_parameters['max_steps'])
gui = GUI.GUI(track, cars, trace=True)
car_objects = [Environment.Car(c) for c in cars]
rl = RL.QLearning_NN(rl_prams, weights_save_dir=weights_save_dir)
rl.generate_nn()
if load_weights is not None:
if load_weights == 'all':
run = True
else:
rl.load_weights(load_weights)
if random_seed is not None: rl.random_seed(random_seed)
weight_names = sorted([name for name in glob.glob(weights_save_dir + '*')])
weight_names_index = 0
def initialize(run_state):
env.compute_interaction(car_objects)
for car in car_objects:
car.reset()
car.get_sensor_reading()
if run_state == True:
env.set_max_steps(1500)
gui.remove_traces()
gui.disable_trace()
gui.set_run_select(gui.runs[1])
gui.update_debug_info('[Testing]\n' +
'Currently learned weights loaded')
else:
env.set_max_steps(rl_prams['max_steps'])
gui.enable_trace()
gui.set_run_select(gui.runs[0])
gui.update_debug_info('[Training]\n')
def check_run_button(current_state):
if gui.get_run_select() == gui.runs[0] and current_state == True:
print '\n\n\nLearning\n'
initialize(run_state=False)
return False
if gui.get_run_select() == gui.runs[1] and run == False:
print '\n\n\nRun only\n'
initialize(run_state=True)
return True
return None
initialize(run_state=run)
while (1):
new_run_state = check_run_button(current_state=run)
if new_run_state is not None: run = new_run_state
if run == True:
for i, car in enumerate(car_objects):
terminal = rl.run_step(car, env, dt)
if terminal is not None:
print 'Car', i, ':', terminal
if i == 0:
if load_weights == 'all' and weight_names_index < len(
weight_names):
rl.load_weights(weight_names[weight_names_index])
gui.update_debug_info(
'[Testing]\n' + 'Weights loaded:\n' +
weight_names[weight_names_index])
weight_names_index += 1
gui.update(i, car.get_state())
env.compute_interaction(car_objects)
gui.refresh()
else:
terminal, debug, epoch, avg_loss, final_score, cross_score = rl.learn_step(
car_objects[0], env, dt)
if terminal is not None:
if debug is not None:
gui.update_debug_info(debug)
gui.update_graph(epoch, avg_loss, gui.graphs[0])
gui.update_graph(epoch, final_score, gui.graphs[1])
gui.update_graph(epoch, cross_score, gui.graphs[2])
gui.refresh()
gui.update(0, terminal, draw_car=False, force_end_line=True)
gui.refresh()
if rl.epoch % 100 == 0:
gui.update(0, car_objects[0].get_state(), draw_car=True)
gui.refresh()
else:
gui.update(0, car_objects[0].get_state(), draw_car=False)
