def __init__(self,
learner,
memory,
simulator,
settings,
dqn_policy=None,
rollout_policy=None):
"""
The learning agent is responsible for communicating and moving
data between the three modules: Learner, Simulator, Memory
Inputs:
- learner: containes the neural network and the optimizer to train it
- memory: expereince replay memory that can be minibatch sampled
- simulator: simulates the environemnt
- settings: hyper parameters for training
- rollout_policy: rollout policy, random by default
"""
self.learner = learner
self.memory = memory
self.simulator = simulator # for populating the experience replay
self.evaluator = deepcopy(simulator) # for evaluation
self.dqn_policy = dqn_policy
if dqn_policy is None:
self.dqn_policy = DQNPolicy(learner)
self.rollout_policy = rollout_policy
if rollout_policy is None:
self.rollout_policy = RandomPolicy(simulator.n_actions)
self.set_params(settings)
self.n_epochs = self.iterations / float(memory.memory_size)
self.iteration = []
self.loss = []
self.q_ave = []
self.eval_iteration = []
self.r_eval = []
self.r_per_episode_eval = []
# run the simulation
input_state = np.zeros((1, 2), dtype=np.float32)
for i in xrange(nsteps):
state = simulator.get_screenshot()
input_state[0] = state
a = policy.action((input_state, None))
simulator.act(a)
r = simulator.reward()
rtot += r
xpos[i], vel[i] = state
if simulator.episode_over():
break
return rtot, xpos, vel
mdp = MountainCar()
simulator = MDPSimulator(mdp)
net = pickle.load(open("../chimp/pre_trained_nets/mountain_car.net", "rb"))
backend = ChainerBackend(settings)
backend.set_net(net)
learner = DQNLearner(settings, backend)
policy = DQNPolicy(learner)
r, xtrace, vtrace = car_sim(300, simulator, policy, verbose=True)
p.plot(xtrace)
p.plot(10.0 * vtrace)
p.show()
