def get_trajectories(batch_size=32,
timesteps=10,
policy='random',
random_start=False,
training=True):
envs = MultiEnvironment([Env() for _ in range(batch_size)])
t_states, t_rewards, t_dones, t_actions = [], [], [], []
# Initial actions/stats
actions = np.random.randint(envs.action_space.n, size=(batch_size, ))
for t in range(timesteps):
states, rewards, dones, _ = envs.step(actions)
rewards = [rewards]
if policy == 'random':
actions = np.random.randint(envs.action_space.n,
size=(batch_size, ))
if policy == 'repeat':
actions = [i % envs.action_space.n for i in range(batch_size)]
t_states.append(states)
t_rewards.append(rewards)
t_dones.append(dones)
t_actions.append(actions)
# Reshape to (batch_size, timesteps, ...)
states = np.swapaxes(t_states, 0, 1)
rewards = np.swapaxes(t_rewards, 0, 1)
dones = np.swapaxes(t_dones, 0, 1)
actions = np.swapaxes(t_actions, 0, 1)
return states, rewards, dones, actions
def get_trajectories(batch_size=32, timesteps=10, policy=None):
envs = MultiEnvironment([BoxesEnv() for _ in range(batch_size)])
t_states, t_rewards, t_dones, t_actions = [], [], [], []
for t in range(timesteps):
actions = np.random.randint(envs.action_space.n, size=(batch_size, ))
states, rewards, dones, _ = envs.step(actions)
t_states.append(states)
t_rewards.append(rewards)
t_dones.append(dones)
t_actions.append(actions)
# Reshape to (batch_size, timesteps, ...)
states = np.swapaxes(t_states, 0, 1)
rewards = np.swapaxes(t_rewards, 0, 1)
dones = np.swapaxes(t_dones, 0, 1)
actions = np.swapaxes(t_actions, 0, 1)
return states, rewards, dones, actions
self.prev_action = np.random.randint(self.num_actions)
def step(self, state):
flip = np.random.random()
if flip > 0.90:
action = 1 # mash the 'shoot' button
elif flip > 0.25:
action = self.prev_action
self.prev_action = action
else:
action = np.random.randint(self.num_actions)
return action
MAX_BATCH_SIZE = 32
envs = MultiEnvironment([GameEnv() for _ in range(MAX_BATCH_SIZE)])
states = envs.reset()
def get_trajectories(batch_size=32, timesteps=10, policy=None, random_start=False):
global states
actions = np.random.randint(envs.action_space.n, size=(batch_size,))
PolicyFn = policy or HeuristicPolicy
policies = [PolicyFn() for _ in range(batch_size)]
t_states, t_rewards, t_dones, t_actions = [], [], [], []
for t in range(timesteps):
actions = [p.step(s) for p, s in zip(policies, states)]
actions = actions[:batch_size] # hack for fixed envs w/ varible batch size
new_states, rewards, dones, _ = envs.step(actions)
t_states.append(new_states)
for i in range(batch_size):