def start_game(game,
state,
directory='tmp',
steps=1000000,
discrete_actions=False,
bk2dir=None):
use_restricted_actions = retro.ACTIONS_FILTERED
if discrete_actions:
use_restricted_actions = retro.ACTIONS_DISCRETE
try:
env = retro.make(game,
state,
scenario='contest',
use_restricted_actions=use_restricted_actions)
except Exception:
env = retro.make(game,
state,
use_restricted_actions=use_restricted_actions)
if bk2dir:
env.auto_record(bk2dir)
env = retro_contest.StochasticFrameSkip(env, n=4, stickprob=0.25)
env = gym.wrappers.TimeLimit(env, max_episode_steps=4500)
env = RemoteEnvWrapper(env, directory)
env.serve(steps)
def make_rand_env(game_states,
stack=2,
scale_rew=True,
color=False,
exp_type=['x'],
exp_const=[0.002],
max_episode_steps=4500,
maml=False,
small_size=False):
"""
Create an environment with some standard wrappers.
"""
game, state = game_states[0]
env = make(game, state)
if maml:
env_rand = RandomEnvironmen2(env, game_states)
else:
env_rand = RandomEnvironmen(env, game_states)
env = retro_contest.StochasticFrameSkip(env_rand, n=4, stickprob=0.25)
env = BackupOriginalData(env)
env = gym.wrappers.TimeLimit(env, max_episode_steps=max_episode_steps)
env = SonicDiscretizer(env)
env = AllowBacktracking(env)
if scale_rew:
env = RewardScaler(env)
env = WarpFrame(env, color, small_size)
assert len(exp_type) == len(exp_const)
for t, c in zip(exp_type, exp_const):
if c > 0:
if t == 'obs':
env = ObsExplorationReward(env, c, game_specific=True)
elif t == 'x':
env = XExplorationReward(env, c, game_specific=True)
else:
raise ValueError('unknown exploration {}'.format(t))
if stack > 1:
env = FrameStack(env, stack)
env = EpisodeInfo(env)
if maml:
env.sample = env_rand.sample
return env
def make(game, state=retro.State.DEFAULT, discrete_actions=False, bk2dir=None):
use_restricted_actions = retro.Actions.FILTERED
if discrete_actions:
use_restricted_actions = retro.Actions.DISCRETE
try:
env = retro.make(game, state, scenario='contest', use_restricted_actions=use_restricted_actions)
except Exception:
env = retro.make(game, state, use_restricted_actions=use_restricted_actions)
if bk2dir:
env.auto_record(bk2dir)
env = retro_contest.StochasticFrameSkip(env, n=4, stickprob=0.25)
env = gym.wrappers.TimeLimit(env, max_episode_steps=4500)
return env
def make(game, state, discrete_actions=False, bk2dir=None, max_episode_steps=4000):
"""Make the competition environment."""
print('game:', game, 'state:', state)
use_restricted_actions = retro.ACTIONS_FILTERED
if discrete_actions:
use_restricted_actions = retro.ACTIONS_DISCRETE
try:
env = retro.make(game, state, scenario='contest', use_restricted_actions=use_restricted_actions)
except Exception:
env = retro.make(game, state, use_restricted_actions=use_restricted_actions)
if bk2dir:
env.auto_record(bk2dir)
env = retro_contest.StochasticFrameSkip(env, n=4, stickprob=0.25)
env = gym.wrappers.TimeLimit(env, max_episode_steps=max_episode_steps)
return env
def make_env(game, state, stack=True, scale_rew=True):
"""
Create an environment with some standard wrappers.
"""
env = make(game, state)
env = retro_contest.StochasticFrameSkip(env, n=4, stickprob=0.25)
env = gym.wrappers.TimeLimit(env, max_episode_steps=4500)
env = SonicDiscretizer(env)
env = AllowBacktracking(env)
if scale_rew:
env = RewardScaler(env)
env = WarpFrame(env)
if stack:
env = FrameStack(env, 4)
env = EpisodeInfo(env)
return env
def make(game, state=retro.STATE_DEFAULT, discrete_actions=False, bk2dir=None,monitordir=None, scenario='contest'):
use_restricted_actions = retro.ACTIONS_FILTERED
if discrete_actions:
use_restricted_actions = retro.ACTIONS_DISCRETE
try:
#env = retro.make(game, state, scenario='contest', use_restricted_actions=use_restricted_actions)
env = retro.make(game, state, scenario=scenario, use_restricted_actions=use_restricted_actions)
except Exception:
env = retro.make(game, state, use_restricted_actions=use_restricted_actions)
if bk2dir:
env.auto_record(bk2dir)
#added this
if monitordir:
env = retro_contest.Monitor(env, os.path.join(monitordir, 'monitor.csv'), os.path.join(monitordir, 'log.csv'))
env = retro_contest.StochasticFrameSkip(env, n=4, stickprob=0.25)
env = gym.wrappers.TimeLimit(env, max_episode_steps=4500)
#env.serve(timestep_limit=10000, ignore_reset=True)
return env
def retro_make(game, state=retro.State.DEFAULT, discrete_actions=False, bk2dir=None, record='.'):
use_restricted_actions = retro.Actions.FILTERED
if discrete_actions:
use_restricted_actions = retro.Actions.DISCRETE
try:
env = retro.make(game, state, scenario='deep_thought', record=record, use_restricted_actions=use_restricted_actions)
except Exception as e:
print('EXCEPTION in retro_make')
print(traceback.format_exc())
env = retro.make(game, state, use_restricted_actions=use_restricted_actions)
if bk2dir:
env.auto_record(bk2dir)
env = retro_contest.StochasticFrameSkip(env, n=4, stickprob=0.25)
if game in game_wrappers:
env = game_wrappers[game](env)
else:
env = gym.wrappers.TimeLimit(env, max_episode_steps=4500)
return env
def make_maml_env(game_states,
stack=2,
scale_rew=True,
color=False,
exp_type='x',
exp_const=0.002,
max_episode_steps=4500):
"""
Create an environment with some standard wrappers.
"""
game, state = game_states[0]
env = make(game, state)
env_rand = RandomEnvironmen2(env, game_states)
env = retro_contest.StochasticFrameSkip(env_rand, n=4, stickprob=0.25)
env = BackupOriginalData(env)
env = gym.wrappers.TimeLimit(env, max_episode_steps=max_episode_steps)
env = SonicDiscretizer(env)
env = AllowBacktracking(env)
if scale_rew:
env = RewardScaler(env)
env = WarpFrame(env, color)
if exp_const > 0:
if exp_type == 'obs':
env = ObsExplorationReward(env, exp_const, game_specific=True)
elif exp_type == 'x':
env = XExplorationReward(env, exp_const, game_specific=True)
if stack > 1:
env = FrameStack(env, stack)
env = ScaledFloatFrame(env)
env = EpisodeInfo(env)
env.sample = env_rand.sample
return env
def make(game,
state=retro.State.DEFAULT,
discrete_actions=False,
bk2dir=None,
monitordir=None,
scenario='scenario'):
use_restricted_actions = retro.Actions.FILTERED #retro.ACTIONS_FILTERED
if discrete_actions:
use_restricted_actions = retro.Actions.DISCRETE #retro.ACTIONS_DISCRETE
try:
#env = retro.make(game, state, scenario='contest', use_restricted_actions=use_restricted_actions)
env = retro.make(game,
state,
scenario=scenario,
use_restricted_actions=use_restricted_actions)
except Exception:
env = retro.make(game,
state,
use_restricted_actions=use_restricted_actions)
if bk2dir:
env.auto_record(bk2dir)
#added this
if monitordir:
time_int = int(time.time())
env = retro_contest.Monitor(
env, os.path.join(monitordir, 'monitor_{}.csv'.format(time_int)),
os.path.join(monitordir, 'log_{}.csv'.format(time_int)))
#bust a move
#env = retro_contest.StochasticFrameSkip(env, n=6, stickprob=0.0) #n=10, did some analysis on this
#contra
env = retro_contest.StochasticFrameSkip(env, n=4, stickprob=0.0)
#sonic
#env = retro_contest.StochasticFrameSkip(env, n=4, stickprob=0.25)
env = gym.wrappers.TimeLimit(env, max_episode_steps=8000)
#env.serve(timestep_limit=10000, ignore_reset=True)
return env