def configure(self, environment_definitions, port_inc=0):
self.action_space = spaces.Discrete(
environment_definitions['action_shape'][0])
min_value = environment_definitions['min_value']
max_value = environment_definitions['max_value']
state_shape = environment_definitions['state_shape']
state_type = environment_definitions['state_type']
self.observation_space = spaces.Box(low=min_value,
high=max_value,
shape=state_shape,
dtype=state_type)
self.n_envs = environment_definitions['n_envs']
self.actions = environment_definitions['actions']
if 'action_meaning' in environment_definitions:
self.action_meaning = environment_definitions['action_meaning']
else:
self.action_meaning = ['action'] * len(self.actions)
if 'agent' in environment_definitions:
if inspect.isclass(environment_definitions['agent']):
self.agent = environment_definitions['agent']()
else:
raise Exception("Agent object is not a class!!!!")
else:
self.agent = BasicAgent()
host = environment_definitions['host']
input_port = environment_definitions['input_port']
output_port = environment_definitions['output_port']
self.remoteenv = RemoteEnv(host, output_port + port_inc,
input_port + port_inc)
self.remoteenv.open(0)
self.nlives = 1
self.configureFlag = True
self.state = None
def agent():
env = RemoteEnv(IN_PORT=8080, OUT_PORT=7070)
env.open(0)
for i in range(10000000):
sum_energy = 0
state = env.step("restart")
prev_energy = state['energy']
print(prev_energy)
done = state['done']
while not done:
frame = image_decode(state['frame'], 20, 20)
print(frame)
action = int(input('action'))
for i in range(8):
state = env.step("act", action)
state = env.step('get_status')
done = state["done"]
energy = state['energy']
frame = state['frame']
touched = state['touched']
ID = state['id']
delta = (energy - prev_energy)
print(delta)
sum_energy += delta
prev_energy = energy
print(sum_energy)
env.close()
def train():
actions = [("fx", 0.1), ("fx", -0.1), ("fz", 0.1), ("fz", -0.1)]
agent = KerasDQNAgent(build_model, action_size=4, decay_steps=20000)
agent.eps_min = 0.001
t = Thread(target=run_worker, args=(agent, ))
t.start()
env = RemoteEnv()
env.open(0)
fields = None
total_steps = 0
for e in range(1000000):
print("Starting new episode ", e)
fields = env.step("restart")
done = fields['done']
reward = 0
initial_state = get_state_from_fields(fields).reshape(1, 7)
steps = 0
sum_rewards = 0
while not done:
if total_steps >= 10000:
agent.training = True
action = agent.predict(initial_state)
else:
action = agent.predict(initial_state, True)
reward = 0.0
for _ in range(4):
fields = env.step(actions[action][0], actions[action][1])
reward += fields['reward']
if fields['done'] == True:
break
final_state = get_state_from_fields(fields).reshape(1, 7)
reward = np.clip(reward, -1.0, 1.0)
sum_rewards += reward
done = fields['done']
if total_steps % 20 == 0:
print("\nLOSS ", agent.last_loss, "\n")
agent.add_sample(initial_state, action, final_state, reward, done)
if steps >= 2000 and not done:
done = True
if not done:
initial_state = np.copy(final_state)
steps += 1
total_steps += 1
print("Sum of rewards ", sum_rewards, ", Steps by episode: ", steps,
", Total Steps: ", total_steps, ", EPS: ", agent.eps,
", TOTAL SAMPLES: ", agent.mem_size())
if e % 50 == 0:
agent.model.save_weights("model.h5", save_format='h5')
agent.done = True
agent.training = False
env.close()
def agent():
env = RemoteEnv(IN_PORT=8081, OUT_PORT=7071)
env.open(0)
for i in range(10000000):
state = env.step("restart")
done = state['done']
print('new episode')
while not done:
frame = image_from_str(state['state'], 10, 10)
print(frame)
action = int(input('action'))
state = env.step("move", action)
done = state['done']
print(state)
env.close()
def run_agent(inport, outport):
env = RemoteEnv(IN_PORT=inport, OUT_PORT=outport)
env.open(0)
env.step("restart")
speed = 100
angular_speed = 50
actions = [('walk', 15), ('run', 30), ('walk_in_circle', 1), ('left_turn', 1), ('right_turn', 1), ('up', 1),
('down', 1), ('jump', True), ('pickup', True), ('pickup', False), ('noop', -1)]
action_size = len(actions)
for i in range(100000):
sum_rewards = 0
touchID = 0
energy = 0
#idx = np.random.choice(len(actions))
idx = int(input())
for i in range(8):
env_info = env.step(actions[idx][0], actions[idx][1])
done = env_info['done']
if done:
break
if not done:
env_info = env.step('get_result', -1)
sum_rewards += env_info['reward']
touchID = env_info['touchID']
energy = env_info['energy']
if done:
sum_rewards += env_info['reward']
touchID = env_info['touchID']
energy = env_info['energy']
env_info = env.step('restart', -1)
print("Object touched ---------------- ", touchID)
print("Reward sum -------------------- ", sum_rewards)
print("Done ---------------------------", env_info['done'])
print("--------------------------------------------------------------------------------------")
print_matrix(get_frame_from_fields(env_info['frame']))
print("-----------------------------------------------")
env.close()
def configure(self, environment_definitions, port_inc=0):
self.action_space = spaces.Discrete(
environment_definitions['action_shape'][0])
min_value = environment_definitions['min_value']
max_value = environment_definitions['max_value']
state_shape = environment_definitions['state_shape']
state_type = environment_definitions['state_type']
self.observation_space = spaces.Box(low=min_value,
high=max_value,
shape=state_shape,
dtype=state_type)
self.n_envs = environment_definitions['n_envs']
self.actions = environment_definitions['actions']
self.action_meaning = environment_definitions['action_meaning']
self.state_wrapper = environment_definitions['state_wrapper']
host = environment_definitions['host']
input_port = environment_definitions['input_port']
output_port = environment_definitions['output_port']
self.remoteenv = RemoteEnv(host, output_port + port_inc,
input_port + port_inc)
self.remoteenv.open(0)
self.nlives = 1
def agent():
env = RemoteEnv(IN_PORT=8085, OUT_PORT=7075)
env.open(0)
for i in range(10000000):
state = env.step("restart")
prev_energy = state['energy']
done = state['done']
while not done:
action = int(input('action'))
#action = np.random.choice([0, 1])
reward_sum = 0
touched = -1
for i in range(8):
state = env.step("act", action)
state = env.step('get_status')
energy = state['energy']
if touched == -1:
touched = state['touched']
elif not state['touched'] in [0, 1]:
touched = state['touched']
if state['done']:
break
done = state['done']
reward_sum += (energy - prev_energy)
prev_energy = energy
frame = state['frame']
ID = state['id']
frame = image_decode(state['frame'], 20, 20)
print(frame)
print(reward_sum)
print(touched)
print(state['signal'])
print(state['done'])
prev_energy = energy
if i >= 2:
break
env.close()
from unityremote.core import RemoteEnv
import numpy as np
env = RemoteEnv()
env.open(0)
if __name__ == "__main__":
for i in range(10):
sum_reward = 0
state = env.step("restart")
print(state)
done = state['done']
while not done:
frame = None
action = int(
input("action-----------------------------------------"))
if action == 0:
state = env.step("fx", 0.1)
elif action == 1:
state = env.step("fz", 0.1)
else:
state = env.step("noop", 0.0)
done = state["done"]
reward = state['reward']
sum_reward += reward
print(sum_reward)
env.close()