def make_rollout_agent(ob_size, action_size, starting_weights, args):
rollout_agent = rollout_master_agent.parallel_rollout_master_agent(
args, ob_size, action_size
)
rollout_agent.set_policy_weights(starting_weights)
return rollout_agent
optimizer_tasks = multiprocessing.JoinableQueue()
optimizer_results = multiprocessing.Queue()
optimizer_agent = optimization_agent.optimization_agent(
args,
learner_env.observation_space.shape[0],
learner_env.action_space.shape[0],
optimizer_tasks,
optimizer_results
)
optimizer_agent.start()
# the rollouts agents
rollout_agent = rollout_master_agent.parallel_rollout_master_agent(
args,
learner_env.observation_space.shape[0],
learner_env.action_space.shape[0]
)
# start the training and rollouting process
optimizer_tasks.put(parallel_util.START_SIGNAL)
optimizer_tasks.join()
starting_weights = optimizer_results.get()
# Note that it is super hacky, if the number of threads is too big,
# when the weights are being assigned, some rollout_agents might
# not even be created.
time.sleep(5)
rollout_agent.set_policy_weights(starting_weights)
else:
# the case where invoking dm_control suite
pass