def __init__(self,
env,
networks,
scope=None,
n_itr=5000,
start_itr=0,
batch_size=32,
max_path_length=200,
discount=0.99,
plot=False,
pause_for_plot=False,
center_adv=True,
max_epsilon=1,
min_epsilon=0.01,
store_paths=False,
whole_paths=True,
sampler_cls=None,
sampler_args=None,
force_batch_sampler=True,
pre_trained_size=10000,
target_network_update=1000,
save_param_update=125,
**kwargs):
"""
:param env: Environment
:param policy: Policy
:type policy: Policy
:param scope: Scope for identifying the algorithm. Must be specified if running multiple algorithms
simultaneously, each using different environments and policies
:param n_itr: Number of iterations.
:param start_itr: Starting iteration.
:param batch_size: Number of samples per iteration.
:param max_path_length: Maximum length of a single rollout.
:param discount: Discount.
:param plot: Plot evaluation run after each iteration.
:param pause_for_plot: Whether to pause before contiuing when plotting.
:param store_paths: Whether to save all paths data to the snapshot.
:return:
"""
self.ma_mode = kwargs.pop('ma_mode', 'centralized')
self.env = env
self.policy = networks['q_network']
self.target_policy = networks['target_q_network']
self.scope = scope
self.n_itr = n_itr
self.start_itr = start_itr
self.batch_size = batch_size
self.max_path_length = max_path_length
self.discount = discount
self.plot = plot
self.max_epsilon = max_epsilon
self.min_epsilon = min_epsilon
self.pause_for_plot = pause_for_plot
self.center_adv = center_adv
self.store_paths = store_paths
self.whole_paths = whole_paths
self.force_batch_sampler = force_batch_sampler
self.loss_after = 0
self.mean_kl = 0
self.max_kl = 0
self.save_param_update = save_param_update
self.target_network_update = target_network_update
self.writer = None
self.write_op = None
self.pre_trained_size = pre_trained_size
self.total_episodic_rewards = None
self.s_loss = None
self.s_avg_rewards = None
self.s_total_rewards = None
if sampler_cls is None:
sampler_cls = ExpReplayMASampler
if sampler_args is None:
sampler_args = dict()
self.sampler = sampler_cls(algo=self, **sampler_args)
if plot:
from rllab.plotter import plotter
plotter.init_worker()
self.init_opt()
def start_worker(self):
self.sampler.start_worker()
if self.plot:
plotter.init_worker()###
plotter.init_plot(self.env, self.policy)