def __init__(self, env):
gym.Wrapper.__init__(self, env)
RewardShapingInterface.__init__(self)
self.num_agents = env.num_agents
self.is_multiagent = env.is_multiagent
# save a reference to this wrapper in the actual env class, for other wrappers and for outside access
self.env.unwrapped.reward_shaping_interface = self
def __init__(self, env, reward_shaping_scheme=None, is_multiagent=False):
gym.Wrapper.__init__(self, env)
RewardShapingInterface.__init__(self)
self.reward_shaping_scheme = reward_shaping_scheme
self.cumulative_rewards = None
self.episode_actions = None
self.num_agents = env.num_agents if hasattr(env, 'num_agents') else 1
self.is_multiagent = is_multiagent
# save a reference to this wrapper in the actual env class, for other wrappers and for outside access
self.env.unwrapped.reward_shaping_interface = self
def __init__(self, env, increase_team_spirit, max_team_spirit_steps):
gym.Wrapper.__init__(self, env)
RewardShapingInterface.__init__(self)
TrainingInfoInterface.__init__(self)
self.num_agents = env.unwrapped.num_agents
self.is_multiagent = env.unwrapped.is_multiagent
self.episode_rewards = [0] * self.num_agents
self.increase_team_spirit = increase_team_spirit
self.max_team_spirit_steps = max_team_spirit_steps
self.approx_total_training_steps = 0
def __init__(self, env, reward_shaping_scheme=None, annealing=None):
gym.Wrapper.__init__(self, env)
RewardShapingInterface.__init__(self)
TrainingInfoInterface.__init__(self)
self.reward_shaping_scheme = reward_shaping_scheme
self.cumulative_rewards = None
self.episode_actions = None
self.num_agents = env.num_agents if hasattr(env, 'num_agents') else 1
self.reward_shaping_updated = True
self.annealing = annealing
def __init__(self, num_agents, make_env_func, env_config, skip_frames):
gym.Env.__init__(self)
RewardShapingInterface.__init__(self)
self.num_agents = num_agents
log.debug('Multi agent env, num agents: %d', self.num_agents)
self.skip_frames = skip_frames # number of frames to skip (1 = no skip)
env = make_env_func(
player_id=-1
) # temporary env just to query observation_space and stuff
self.action_space = env.action_space
self.observation_space = env.observation_space
self.default_reward_shaping = get_default_reward_shaping(env)
env.close()
self.current_reward_shaping = [
self.default_reward_shaping for _ in self.num_agents
]
self.make_env_func = make_env_func
self.safe_init = env_config is not None and env_config.get(
'safe_init', False)
if self.safe_init:
sleep_seconds = env_config.worker_index * 1.0
log.info(
'Sleeping %.3f seconds to avoid creating all envs at once',
sleep_seconds)
time.sleep(sleep_seconds)
log.info('Done sleeping at %d', env_config.worker_index)
self.env_config = env_config
self.workers = None
# only needed when rendering
self.enable_rendering = False
self.last_obs = None
self.reset_on_init = True
self.initialized = False
def __init__(self, env, reward_shaping_scheme=None, true_reward_func=None):
gym.Wrapper.__init__(self, env)
RewardShapingInterface.__init__(self)
self.reward_shaping_scheme = reward_shaping_scheme
self.true_reward_func = true_reward_func
# without this we reward using BFG and shotguns too much
self.reward_delta_limits = dict(DAMAGECOUNT=200, HITCOUNT=5)
self.prev_vars = dict()
self.prev_dead = True
self.orig_env_reward = self.total_shaping_reward = 0.0
self.selected_weapon = deque([], maxlen=5)
self.reward_structure = {}
self.verbose = False
self.print_once = False
# save a reference to this wrapper in the actual env class, for other wrappers
self.env.unwrapped.reward_shaping_interface = self