tags={'wrapper_config.TimeLimit.max_episode_steps': 5000},
reward_threshold=4750.0,
)
env = gym.make(ENV)
input_shape = env.observation_space.shape[0]
output_shape = env.action_space.n
print('observation dim:', (input_shape, ))
print('action dim: ', (output_shape, ))
#logging/debug
info = Info(SummaryWriter(comment="-cartpole-pg"))
ag = ReinforceAgent(PolicyGradientNetwork(input_shape, output_shape))
sim = pwsim.GymSimulator(env, info)
sim.add_agent(ag)
print('Training: ', ENV)
for t in sim:
avg_reward = info.info[info.info_labels[0]]
if t.episode > TIMEOUT or avg_reward > 2000:
break
########## TEST and render
print("TEST!")
#env = gym.make('CartPole-long-v0')
env = gym.wrappers.Monitor(env, './videos', force=True)
@author: ben
"""
from pyworld import agent as pwag
from pyworld import simulate as pwsim
class TestAgent(pwag.Agent):
def __init__(self, sensors, actuators):
super(TestAgent, self).__init__(sensors, actuators)
sensors[0]._callback = self.sense
sensors[1]._callback = self.sense
def sense(self, obs):
print(obs)
def attempt(self, state):
self.actuators[0]()
sim = pwsim.GymSimulator('CartPole-v0')
ag = TestAgent(
[pwag.EpisodicSensor(), pwag.SimpleSensor()],
[pwag.RandomActuator(sim.env.action_space)])
sim.add_agent(ag)
for t in sim:
if t.done:
sim.stop()
print(t)
REWARD_STEPS = 6 # number of steps to look in future (discounted reward)
GAMMA = 0.99 # reward discount (for infinite horizon problems)
SKIP = 3 # number of frames to skip at each step (frames come in too fast, take the nth frame as the observation)
STACK = 4 # number of frames to stack. helps make the state space markovian
EPSILON_START = 1.0 # epsilon for e-greedy policy
EPSILON_END = 0.02
if (torch.cuda.is_available()
): # use cuda (GPU compute for faster training!)
print("USING CUDA!")
DEVICE = 'cuda'
else:
DEVICE = 'cpu'
sim = pwsim.GymSimulator('Breakout-v0')
action_shape = sim.env.action_space.n
observation_shape = [STACK, 84, 84] #todo find a way to do this nicely
ag = DQNAgent(DEVICE, GAMMA)
epsilon_tracker = pwag.EpsilonTracker(epsilon_start=EPSILON_START,
epsilon_end=EPSILON_END)
actuator = pwag.EpsilonGreedyActuator(epsilon_tracker)
sensor = pwag.MaxPoolSensor(pwag.AtariImageSensor(pwag.BufferedSensor(ag)))
ag.add_component('sensor', sensor)
ag.add_component('actuator', actuator)
ag.add_component('experience_replay', pwag.ExperienceReplay())