def mp__update_params(args, pipe1_eva, pipe1_exp_list):
args.init_before_training(if_main=False)
'''basic arguments'''
env = args.env
cwd = args.cwd
agent = args.agent
rollout_num = args.rollout_num
'''training arguments'''
net_dim = args.net_dim
max_memo = args.max_memo
break_step = args.break_step
batch_size = args.batch_size
target_step = args.target_step
repeat_times = args.repeat_times
if_break_early = args.if_break_early
del args # In order to show these hyper-parameters clearly, I put them above.
'''init: environment'''
max_step = env.max_step
state_dim = env.state_dim
action_dim = env.action_dim
if_discrete = env.if_discrete
'''init: Agent, ReplayBuffer'''
agent.init(net_dim, state_dim, action_dim)
if_on_policy = agent.__class__.__name__ in {'AgentPPO', 'AgentGaePPO', 'AgentInterPPO'}
'''send'''
pipe1_eva.send(agent.act) # send
# act = pipe2_eva.recv() # recv
buffer_mp = ReplayBufferMP(max_len=max_memo + max_step * rollout_num, if_on_policy=if_on_policy,
state_dim=state_dim, action_dim=1 if if_discrete else action_dim,
rollout_num=rollout_num, if_gpu=True)
'''prepare for training'''
if if_on_policy:
steps = 0
else: # explore_before_training for off-policy
with torch.no_grad(): # update replay buffer
steps = 0
for i in range(rollout_num):
pipe1_exp = pipe1_exp_list[i]
# pipe2_exp.send((buffer.buf_state[:buffer.now_len], buffer.buf_other[:buffer.now_len]))
buf_state, buf_other = pipe1_exp.recv()
steps += len(buf_state)
buffer_mp.extend_buffer(buf_state, buf_other, i)
agent.update_net(buffer_mp, target_step, batch_size, repeat_times) # pre-training and hard update
agent.act_target.load_state_dict(agent.act.state_dict()) if getattr(env, 'act_target', None) else None
agent.cri_target.load_state_dict(agent.cri.state_dict()) if getattr(env, 'cri_target', None) in dir(
agent) else None
total_step = steps
'''send'''
pipe1_eva.send((agent.act, steps, 0, 0.5)) # send
# act, steps, obj_a, obj_c = pipe2_eva.recv() # recv
'''start training'''
if_solve = False
while not ((if_break_early and if_solve)
or total_step > break_step
or os.path.exists(f'{cwd}/stop')):
'''update ReplayBuffer'''
steps = 0 # send by pipe1_eva
for i in range(rollout_num):
pipe1_exp = pipe1_exp_list[i]
'''send'''
pipe1_exp.send(agent.act)
# agent.act = pipe2_exp.recv()
'''recv'''
# pipe2_exp.send((buffer.buf_state[:buffer.now_len], buffer.buf_other[:buffer.now_len]))
buf_state, buf_other = pipe1_exp.recv()
steps += len(buf_state)
buffer_mp.extend_buffer(buf_state, buf_other, i)
total_step += steps
'''update network parameters'''
obj_a, obj_c = agent.update_net(buffer_mp, target_step, batch_size, repeat_times)
'''saves the agent with max reward'''
'''send'''
pipe1_eva.send((agent.act, steps, obj_a, obj_c))
# q_i_eva_get = pipe2_eva.recv()
if_solve = pipe1_eva.recv()
if pipe1_eva.poll():
'''recv'''
# pipe2_eva.send(if_solve)
if_solve = pipe1_eva.recv()
buffer_mp.print_state_norm(env.neg_state_avg if hasattr(env, 'neg_state_avg') else None,
env.div_state_std if hasattr(env, 'div_state_std') else None) # 2020-12-12
'''send'''
pipe1_eva.send('stop')
# q_i_eva_get = pipe2_eva.recv()
time.sleep(4)
def mp__update_params(args, pipe1_eva, pipe1_exp_list):
agent_rl = args.agent_rl # basic arguments
env = args.env
cwd = args.cwd
rollout_num = args.rollout_num
gamma = args.gamma # training arguments
net_dim = args.net_dim
max_memo = args.max_memo
target_step = args.target_step
batch_size = args.batch_size
repeat_times = args.repeat_times
reward_scale = args.reward_scale
break_step = args.break_step
if_break_early = args.if_break_early
del args # In order to show these hyper-parameters clearly, I put them above.
'''init: env'''
state_dim = env.state_dim
action_dim = env.action_dim
if_discrete = env.if_discrete
max_step = env.max_step
'''build agent'''
agent = agent_rl(net_dim, state_dim, action_dim) # build AgentRL
pipe1_eva.send(agent.act) # act = pipe2_eva.recv()
if_on_policy = agent_rl.__name__ in {'AgentPPO', 'AgentGaePPO'}
buffer_mp = ReplayBufferMP(max_memo + max_step * rollout_num, state_dim,
if_on_policy=if_on_policy,
action_dim=1 if if_discrete else action_dim,
rollout_num=rollout_num) # build experience replay buffer
steps = 0
if not if_on_policy:
with torch.no_grad(): # update replay buffer
for _buffer in buffer_mp.buffers:
steps += _explore_before_train(env, _buffer, target_step // rollout_num, reward_scale, gamma)
agent.update_net(buffer_mp, target_step, batch_size, repeat_times) # pre-training and hard update
agent.act_target.load_state_dict(agent.act.state_dict()) if 'act_target' in dir(agent) else None
total_step = steps
pipe1_eva.send((agent.act, steps, 0, 0.5)) # pipe1_eva (act, steps, obj_a, obj_c)
if_solve = False
while not ((if_break_early and if_solve)
or total_step > break_step
or os.path.exists(f'{cwd}/stop')):
'''update ReplayBuffer'''
for i in range(rollout_num):
pipe1_exp = pipe1_exp_list[i]
pipe1_exp.send(agent.act)
# agent.act = pipe2_exp.recv()
# pipe2_exp.send((buffer.buf_state[:buffer.now_len], buffer.buf_other[:buffer.now_len]))
buf_state, buf_other = pipe1_exp.recv()
steps = len(buf_state)
total_step += steps
buffer_mp.extend_memo_mp(buf_state, buf_other, i)
'''update network parameters'''
obj_a, obj_c = agent.update_net(buffer_mp, target_step, batch_size, repeat_times)
'''saves the agent with max reward'''
pipe1_eva.send((agent.act, steps, obj_a, obj_c)) # pipe1_eva act_cpu
if_solve = pipe1_eva.recv()
if pipe1_eva.poll():
if_solve = pipe1_eva.recv() # pipe2_eva.send(if_solve)
buffer_mp.print_state_norm(env.neg_state_avg if hasattr(env, 'neg_state_avg') else None,
env.div_state_std if hasattr(env, 'div_state_std') else None) # 2020-12-12
pipe1_eva.send('stop') # eva_pipe stop # send to mp_evaluate_agent
time.sleep(4)