def main(args):
dir_name = ROLLOUT_DIR + "/rollout_" + args.env_name + "/"
episodes = args.N
action_dim = action_space_dimension(args.env_name)
epochs = args.epochs
time_steps = args.time_steps
[next_states, correct_state] = import_data(episodes,action_dim,dir_name, time_steps)
try:
agent = StateAgent(action_dim,args.env_name)
except:
print('NO DATA FOUND')
raise
agent.train(next_states,correct_state,epochs)
agent.save_weights()
def test_against_environment(env_name, num_runs, agent_name):
env = gym.make(env_name)
# env.seed(0)
try:
predictor = load_predictive_model(env_name, env.action_space.n)
if agent_name == 'Next_agent':
agent = StateAgent(env.action_space.n, env_name)
agent.set_weights()
elif agent_name == 'DQN':
agent = Agent(gamma=0.99,
epsilon=0.00,
alpha=0.0001,
input_dims=(104, 80, 4),
n_actions=env.action_space.n,
mem_size=25000,
eps_min=0.00,
batch_size=32,
replace=1000,
eps_dec=1e-5,
env_name=env_name)
agent.load_models()
except:
print(
"Error loading model, check environment name and action space dimensions"
)
rewards = []
start = time.time()
total_steps = 0.0
for i in range(num_runs):
frame_queue = deque(maxlen=4)
observation = env.reset()
done = False
if agent_name == 'DQN':
init_queue(frame_queue, observation, True)
else:
init_queue(frame_queue, observation)
total_reward = 0.0
frame_count = 0
while not done:
observation_states = np.concatenate(frame_queue, axis=2)
# Human start of breakout since the next state agent just keeps moving to the left
if agent_name == 'Next_agent':
if env_name == 'BreakoutDeterministic-v4' and not frame_count:
agent_action = 1
else:
next_states = predictor.generate_output_states(
np.expand_dims(observation_states, axis=0))
agent_action = agent.choose_action_from_next_states(
np.expand_dims(next_states, axis=0))
elif agent_name == 'DQN':
agent_action = agent.choose_action(observation_states)
else:
agent_action = env.action_space.sample()
observation, reward, done, _ = env.step(agent_action)
total_reward += reward
frame_count += 1
total_steps += 1
frame_queue.pop()
if agent_name == 'DQN':
frame_queue.appendleft(preprocess_frame_dqn(observation))
else:
frame_queue.appendleft(preprocess_frame(observation))
print("Completed episode {} with reward {}".format(
i + 1, total_reward))
rewards.append(total_reward)
end = time.time()
time_taken = (end - start) / total_steps
print("Test complete - Average score: {} Max score: {}".format(
np.average(rewards), np.max(rewards)))
return (rewards, time_taken)