next_state, reward, done, _ = env.step(env_action)
agent.memory.add(state, action, reward, next_state, float(done))
state = next_state
episode_rewards[-1] += reward
if done:
state = env.reset()
episode_rewards.append(0.0)
if t > hyper_params[
'learning_starts'] and t % hyper_params['learning_freq'] == 0:
agent.optimise_td_loss()
if t > hyper_params['learning_starts'] and t % hyper_params[
'target_update_freq'] == 0:
agent.update_target_network()
num_episodes = len(episode_rewards)
if done and hyper_params['print_freq'] is not None and len(
episode_rewards) % hyper_params['print_freq'] == 0:
mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1)
print("********************************************************")
print("steps: {}".format(t))
print("episodes: {}".format(num_episodes))
print("mean 100 episode reward: {}".format(mean_100ep_reward))
print("% time spent exploring: {}".format(int(100 *
eps_threshold)))
print("********************************************************")
im2vid(
run_episode(agent.policy_network, env, policy_actions, 200,
def main():
config = {'starting-floor': 0, 'total-floors': 9, 'dense-reward': 1,
'lighting-type': 0, 'visual-theme': 0, 'default-theme': 0, 'agent-perspective': 1, 'allowed-rooms': 0,
'allowed-modules': 0,
'allowed-floors': 0,
}
env = ObstacleTowerEnv('./ObstacleTower/obstacletower',
worker_id=1, retro=True, realtime_mode=False, config=config)
print(env.observation_space)
print(env.action_space)
hyper_params = {
"seed": 6, # which seed to use
"replay-buffer-size": int(5e3), # replay buffer size
"learning-rate": 1e-4, # learning rate for Adam optimizer
"discount-factor": 0.99, # discount factor
"num-steps": int(1e6), # total number of steps to run the environment for
"batch-size": 32, # number of transitions to optimize at the same time
"learning-starts": 5000, # number of steps before learning starts
"learning-freq": 1, # number of iterations between every optimization step
"use-double-dqn": True, # use double deep Q-learning
"target-update-freq": 1000, # number of iterations between every target network update
"eps-start": 1.0, # e-greedy start threshold
"eps-end": 0.01, # e-greedy end threshold
"eps-fraction": 0.05, # fraction of num-steps
"print-freq": 10
}
np.random.seed(hyper_params["seed"])
random.seed(hyper_params["seed"])
#assert "NoFrameskip" in hyper_params["env"], "Require environment with no frameskip"
#env = gym.make(hyper_params["env"])
env.seed(hyper_params["seed"])
#env = NoopResetEnv(env, noop_max=30)
#env = MaxAndSkipEnv(env, skip=4)
#env = EpisodicLifeEnv(env)
#env = FireResetEnv(env)
# env = WarpFrame(env)
env = PyTorchFrame(env)
# env = ClipRewardEnv(env)
# env = FrameStack(env, 4)
replay_buffer = ReplayBuffer(hyper_params["replay-buffer-size"])
agent = DQNAgent(
env.observation_space,
env.action_space,
replay_buffer,
use_double_dqn=hyper_params["use-double-dqn"],
lr=hyper_params["learning-rate"],
batch_size=hyper_params["batch-size"],
gamma=hyper_params["discount-factor"]
)
model_num = 500
agent.policy_network.load_state_dict(torch.load('./Models/' + str(model_num) + '_policy.pt',map_location=torch.device(device)))
eps_timesteps = hyper_params["eps-fraction"] * float(hyper_params["num-steps"])
episode_rewards = [0.0]
ep_nums = model_num
state = env.reset()
for t in range(hyper_params["num-steps"]):
fraction = min(1.0, float(t) / eps_timesteps)
eps_threshold = hyper_params["eps-start"] + fraction * (hyper_params["eps-end"] - hyper_params["eps-start"])
sample = random.random()
# TODO
# select random action if sample is less equal than eps_threshold
# take step in env
# add state, action, reward, next_state, float(done) to reply memory - cast done to float
# add reward to episode_reward
if sample > eps_threshold:
action = agent.act(np.array(state))
else:
action = env.action_space.sample()
next_state, reward, done, _ = env.step(action)
agent.memory.add(state, action, reward, next_state, float(done))
state = next_state
episode_rewards[-1] += reward
if done:
state = env.reset()
episode_rewards.append(0.0)
ep_nums += 1
if ep_nums % 50 == 0:
agent.save_models(ep_nums)
plot(episode_rewards,ep_nums)
if t > hyper_params["learning-starts"] and t % hyper_params["learning-freq"] == 0:
agent.optimise_td_loss()
if t > hyper_params["learning-starts"] and t % hyper_params["target-update-freq"] == 0:
agent.update_target_network()
num_episodes = len(episode_rewards)
if done and hyper_params["print-freq"] is not None and len(episode_rewards) % hyper_params[
"print-freq"] == 0:
mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1)
print("********************************************************")
print("steps: {}".format(t))
print("episodes: {}".format(num_episodes))
print("mean 100 episode reward: {}".format(mean_100ep_reward))
print("% time spent exploring: {}".format(int(100 * eps_threshold)))
print("********************************************************")
#if done and ep_nums % 10 == 0:
# animate(env,agent,"anim/progress_"+str(ep_nums))
# state = env.reset()
animate(env,agent,"anim/final")
env.close()