vertical_edge_mat = np.zeros([28,28])
vertical_edge_mat[:,14:] = 1.0
recorder = Recorder(n=6)
debu2el = np.diag(np.ones([10-1]),k=1)+np.eye(10)
# debu2el = debu2el[:-1,:]
scene = syc.Scene(image_matrix=vertical_edge_mat)
sensor = syc.Sensor()
agent = syc.Agent(max_q = [scene.maxx-sensor.hp.winx,scene.maxy-sensor.hp.winy])
reward = syc.Rewards()
RL = DeepQNetwork(len(agent.hp.action_space), sensor.hp.winx+2,#sensor.frame_size+2,
reward_decay=0.9,
e_greedy=0.99,
e_greedy0=0.25,
replace_target_iter=10,
memory_size=30000,
e_greedy_increment=0.001,
state_table=None
)
hp.scene = scene.hp
hp.sensor = sensor.hp
hp.agent = agent.hp
hp.reward = reward.hp
hp.RL = RL.hp
run_env()
scene = syc.Scene(frame_list=images)
sensor = syc.Sensor( log_mode=False, log_floor = 1.0)
agent = syc.Agent(max_q = [scene.maxx-sensor.hp.winx,scene.maxy-sensor.hp.winy])
reward = syc.Rewards(reward_types=['central_rms_intensity', 'speed', 'saccade'],relative_weights=[1.0,-float(sys.argv[1]),-200])
# observation_size = sensor.hp.winx*sensor.hp.winy*2
observation_size = 64*64
RL = DeepQNetwork(len(agent.hp.action_space), observation_size,
n_features_shaped=list(np.shape(sensor.dvs_view))+[1],
shape_fun=lambda x: np.reshape(x,[-1]+list(np.shape(sensor.dvs_view))+[1]),
reward_decay=0.99,
e_greedy=0.95,
e_greedy0=0.8,
replace_target_iter=10,
memory_size=100000,
e_greedy_increment=0.0001,
learning_rate=0.0025,
double_q=True,
dqn_mode=True,
state_table=np.zeros([1,observation_size*hp.mem_depth]),
soft_q_type='boltzmann',
beta=0.1,
arch='conv'
)
# RL.dqn.load_nwk_param('tempX_1.nwk')
# RL.dqn.save_nwk_param('liron_encircle.nwk')
hp.scene = scene.hp
hp.sensor = sensor.hp
hp.agent = agent.hp
hp.reward = reward.hp
hp.RL = RL.hp
recorder = Recorder(n=5)
scene = syc.Scene(image_matrix=np.array(images[0]).reshape([28, 28]))
sensor = syc.Sensor()
agent = syc.Agent(
max_q=[scene.maxx - sensor.hp.winx, scene.maxy - sensor.hp.winy])
reward = syc.Rewards()
observation_size = sensor.hp.winx * sensor.hp.winy
RL = DeepQNetwork(
len(agent.hp.action_space),
observation_size * hp.mem_depth, #sensor.frame_size+2,
reward_decay=0.99,
e_greedy=0.9,
e_greedy0=0.8,
replace_target_iter=10,
memory_size=100000,
e_greedy_increment=0.0001,
learning_rate=0.0025,
double_q=False,
dqn_mode=True,
state_table=np.zeros([1, observation_size * hp.mem_depth]))
hp.scene = scene.hp
hp.sensor = sensor.hp
hp.agent = agent.hp
hp.reward = reward.hp
hp.RL = RL.hp
with open(hp_file, 'wb') as f:
pickle.dump(hp, f)
run_env()
# end of game
print('game over')
#env.destroy()
if __name__ == "__main__":
# maze game
env = Image_env1(bmp_features=BMP_MODE)
all_observations_for_mapping = env.observation_space(
) if not BMP_MODE else None
RL = DeepQNetwork(
env.n_actions,
env.n_features,
#learning_rate=0.00005,
reward_decay=0.5,
e_greedy=0.8,
e_greedy0=0.5,
replace_target_iter=20,
memory_size=300000,
# memory_size=1000,
e_greedy_increment=0.001,
# output_graph=True
state_table=all_observations_for_mapping)
run_img()
# env.after(100, run_img)
# print('-----------------------------------------')
# env.mainloop()
env.plot_reward()
env.save_train_history()
#RL.plot_cost()