def get_random_state(game_name, random_steps=20):
gym_env = gym.make(game_name)
env = CustomGym(gym_env, 'SpaceInvaders-v0')
# Reset the environment and then take 20 random actions
state = env.reset()
for _ in range(random_steps):
state, _, _, _ = env.step(random.randrange(env.action_size))
return state
def play(agent,
game_name,
render=True,
num_episodes=10,
fps=5.0,
monitor=True):
gym_env = gym.make(game_name)
if monitor:
print(gym_env)
gym_env = wrappers.Monitor(gym_env, 'videos/-v0')
print(gym_env)
print(game_name)
env = CustomGym(game_name)
desired_frame_length = 1.0 / fps
episode_rewards = []
episode_vals = []
t = 0
for ep in range(num_episodes):
print("Starting episode", ep)
episode_reward = 0
state = env.reset()
terminal = False
current_time = time()
while not terminal:
policy, value = agent.get_policy_and_value(state)
action_idx = np.random.choice(agent.action_size, p=policy)
state, reward, terminal, _ = env.step(action_idx)
if render:
env.render()
t += 1
episode_vals.append(value)
episode_reward += reward
# Sleep so the frame rate is correct
next_time = time()
frame_length = next_time - current_time
if frame_length < desired_frame_length:
sleep(desired_frame_length - frame_length)
current_time = next_time
episode_rewards.append(episode_reward)
if monitor:
gym_env.monitor.close()
return episode_rewards, episode_vals