print('Collecting experience...')
# episode step for accumulate reward
epinfobuf = deque(maxlen=100)
# check learning time
start_time = time.time()
# env reset
s = np.array(env.reset())
# Trainning
for step in range(1, STEP_NUM // N_ENVS + 1):
a = quota.choose_action(s, EPSILON, EPSILON_O)
# take action and get next state
s_, r, done, infos = env.step(a)
# log arrange
for info in infos:
maybeepinfo = info.get('episode')
if maybeepinfo: epinfobuf.append(maybeepinfo)
s_ = np.array(s_)
# clip rewards for numerical stability
clip_r = np.sign(r)
# store the transition
for i in range(N_ENVS):
quota.store_transition(s[i], a[i], clip_r[i], s_[i], done[i],
quota.options[i].item())
# annealing the epsilon(exploration strategy)
def train(args):
print(args)
torch.manual_seed(args.seed)
np.random.seed(args.seed)
if args.return_function == "GAE":
return_function = GAE
elif args.return_function == "Q":
return_function = Q
elif args.return_function == "A":
return_function = A
MONTE_CARLO = True if args.num_steps == 200 else False
envs = SubprocVecEnv(
[make_env(args.env, i + args.num_envs) for i in range(args.num_envs)],
MONTE_CARLO)
test_env = gym.make(args.env)
test_env.seed(args.seed + args.num_envs)
policy = ActorCriticMLP(input_dim=envs.observation_space.shape[0],
n_acts=envs.action_space.n)
optim = torch.optim.Adam(params=policy.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
test_rewards = []
steps = 1
obs = torch.from_numpy(envs.reset())
while steps < args.max_steps:
logp_actions = []
state_values = []
rewards = []
masks = []
for _ in range(args.num_steps):
probs, state_value = policy.forward(obs)
dist = Categorical(probs)
action = dist.sample()
obs, reward, done, _ = envs.step(action.numpy())
logp_actions.append(dist.log_prob(action).unsqueeze(1))
state_values.append(state_value)
rewards.append(torch.FloatTensor(reward).unsqueeze(1))
masks.append(torch.FloatTensor(1 - done).unsqueeze(1))
obs = torch.from_numpy(obs)
steps += 1
if steps % args.test_every == 0:
test_reward = np.mean(
[test(test_env, policy) for _ in range(10)])
test_rewards.append(test_reward)
print(f"Running reward at timestep {steps}: and {test_reward}")
if (1 - done).sum() == 0:
break
next_value = 0
if not (1 - done).sum() == 0:
_, next_value = policy(obs)
returns = return_function(next_value, rewards, masks, state_values,
args)
loss = policy_gradient(logp_actions, returns)
optim.zero_grad()
loss.backward()
optim.step()
# if monte carlo, we need to reset the environment by hand
if MONTE_CARLO:
obs = torch.from_numpy(envs.reset())
return test_rewards
