def demo_continuous_action_off_policy():
args = Arguments()
args.gpu_id = sys.argv[-1][-4]
from elegantrl2.tutorial.agent import AgentSAC # AgentDDPG AgentTD3
args.agent = AgentSAC()
'''choose environment'''
if_train_pendulum = 1
if if_train_pendulum:
"TotalStep: 4e5, TargetReward: -200, UsedTime: 400s"
env = gym.make('Pendulum-v0')
env.target_return = -200 # set target_reward manually for env 'Pendulum-v0'
args.env = PreprocessEnv(env=env)
args.reward_scale = 2**-3 # RewardRange: -1800 < -200 < -50 < 0
args.net_dim = 2**7
args.batch_size = args.net_dim
args.target_step = args.env.max_step * 4
if_train_lunar_lander = 0
if if_train_lunar_lander:
"TotalStep: 4e5, TargetReward: 200, UsedTime: 900s"
args.env = PreprocessEnv(env=gym.make('LunarLanderContinuous-v2'))
args.reward_scale = 2**0 # RewardRange: -800 < -200 < 200 < 302
if_train_bipedal_walker = 0
if if_train_bipedal_walker:
"TotalStep: 8e5, TargetReward: 300, UsedTime: 1800s"
args.env = PreprocessEnv(env=gym.make('BipedalWalker-v3'))
args.reward_scale = 2**0 # RewardRange: -200 < -150 < 300 < 334
args.gamma = 0.97
args.if_per_or_gae = True
'''train and evaluate'''
train_and_evaluate(args)
def demo_discrete_action_on_policy():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
from elegantrl2.tutorial.agent import AgentDiscretePPO
args.agent = AgentDiscretePPO()
'''choose environment'''
if_train_cart_pole = 0
if if_train_cart_pole:
"TotalStep: 5e4, TargetReward: 200, UsedTime: 60s"
args.env = PreprocessEnv(env='CartPole-v0')
args.net_dim = 2**7
args.batch_size = args.net_dim * 2
args.repeat_times = 2**4
args.target_step = args.env.max_step * 8
args.if_per_or_gae = True
if_train_lunar_lander = 1
if if_train_lunar_lander:
"TotalStep: 2e5, TargetReturn: 200, UsedTime: 400s, LunarLander-v2, PPO"
args.env = PreprocessEnv(env=gym.make('LunarLander-v2'))
args.agent.cri_target = False
args.reward_scale = 2**-1
args.net_dim = 2**8
args.batch_size = args.net_dim * 4
args.target_step = args.env.max_step * 4
args.repeat_times = 2**5
args.if_per_or_gae = True
'''train and evaluate'''
train_and_evaluate(args)
def demo_continuous_action_on_policy():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
from elegantrl2.tutorial.agent import AgentPPO
args.agent = AgentPPO()
'''choose environment'''
if_train_pendulum = 0
if if_train_pendulum:
"TotalStep: 4e5, TargetReward: -200, UsedTime: 400s"
env = gym.make('Pendulum-v0')
env.target_return = -200 # set target_reward manually for env 'Pendulum-v0'
args.env = PreprocessEnv(env=env)
args.reward_scale = 2**-3 # RewardRange: -1800 < -200 < -50 < 0
args.net_dim = 2**7
args.batch_size = args.net_dim * 2
args.target_step = args.env.max_step * 16
if_train_lunar_lander = 0
if if_train_lunar_lander:
"TotalStep: 4e5, TargetReward: 200, UsedTime: 900s"
args.env = PreprocessEnv(env=gym.make('LunarLanderContinuous-v2'))
args.reward_scale = 2**0 # RewardRange: -800 < -200 < 200 < 302
if_train_bipedal_walker = 1
if if_train_bipedal_walker:
"TotalStep: 8e5, TargetReward: 300, UsedTime: 1800s"
args.env = PreprocessEnv(env=gym.make('BipedalWalker-v3'))
args.reward_scale = 2**0 # RewardRange: -200 < -150 < 300 < 334
args.gamma = 0.97
args.if_per_or_gae = True
# args.agent.lambda_entropy = 0.05
'''train and evaluate'''
train_and_evaluate(args)
def demo_discrete_action_off_policy():
args = Arguments()
from elegantrl2.tutorial.agent import AgentDoubleDQN # AgentDQN
args.agent = AgentDoubleDQN()
'''choose environment'''
if_train_cart_pole = 0
if if_train_cart_pole:
"TotalStep: 5e4, TargetReward: 200, UsedTime: 60s"
args.env = PreprocessEnv(env='CartPole-v0')
args.net_dim = 2**7
args.target_step = args.env.max_step * 2
if_train_lunar_lander = 1
if if_train_lunar_lander:
"TotalStep: 2e5, TargetReturn: 200, UsedTime: 400s, LunarLander-v2, PPO"
args.env = PreprocessEnv(env=gym.make('LunarLander-v2'))
args.net_dim = 2**8
args.batch_size = args.net_dim
'''train and evaluate'''
train_and_evaluate(args)
def demo_discrete_action_off_policy():
args = Arguments()
from elegantrl2.agent import AgentD3QN as Agent # AgentDQN
args.agent = Agent()
'''choose environment'''
if_train_cart_pole = 0
if if_train_cart_pole:
"TotalStep: 5e4, TargetReward: 200, UsedTime: 60s"
args.env = PreprocessEnv(env='CartPole-v0')
args.net_dim = 2**7
args.target_step = args.env.max_step * 2
train_and_evaluate(args)
if_train_lunar_lander = 1
if if_train_lunar_lander:
"TotalStep: 2e5, TargetReturn: 200, UsedTime: 400s, LunarLander-v2, PPO"
args.env = PreprocessEnv(env=gym.make('LunarLander-v2'))
args.net_dim = 2**8
args.batch_size = args.net_dim
train_and_evaluate(args)
if_train_lunar_lander_mp = 0
if if_train_lunar_lander_mp:
"TotalStep: 36e4, EpisodeReturn: 200, UsedTime: 1181s, LunarLander-v2, D3QN"
"TotalStep: 52e4, EpisodeReturn: 260, UsedTime: 2148s, LunarLander-v2, D3QN"
args.env = PreprocessEnv(env=gym.make('LunarLander-v2'))
args.net_dim = 2**8
args.batch_size = args.net_dim
args.target_step = args.env.max_step * 1
args.worker_num = 4
args.reward_scale = 2**-4
args.max_memo = 2**19
args.break_step = 2**19
args.if_allow_break = False
args.gpu_id = sys.argv[-1]
train_and_evaluate_mp(args)
def demo_custom_env_finance_rl_nas89(): # 1.7+ 2.0+
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
args.random_seed = 19430
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.lambda_entropy = 0.02
from envs.FinRL.StockTrading import StockEnvNAS89
args.gamma = 0.999
args.env = StockEnvNAS89(if_eval=False,
gamma=args.gamma,
turbulence_thresh=30)
args.eval_env = StockEnvNAS89(if_eval=True,
gamma=args.gamma,
turbulence_thresh=15)
args.net_dim = 2**9
args.repeat_times = 2**4
args.learning_rate = 2**-14
args.batch_size = args.net_dim * 4
args.eval_gap = 2**8
args.eval_times1 = 2**0
args.eval_times2 = 2**1
args.break_step = int(8e6)
args.if_allow_break = False
if_single_proc = 0
if if_single_proc:
args.gpu_id = int(sys.argv[-1][-4])
args.random_seed += int(args.gpu_id)
args.target_step = args.env.max_step * 4
train_and_evaluate(args)
if_single_env = 1
if if_single_env:
args.gpu_id = int(sys.argv[-1][-4])
args.random_seed += int(args.gpu_id)
args.target_step = args.env.max_step * 1
args.worker_num = 4
train_and_evaluate_mp(args)
if_multi_learner = 0
if if_multi_learner:
args.gpu_id = (2, 3) if len(sys.argv) == 1 else eval(
sys.argv[-1]) # python main.py -GPU 0,1
args.repeat_times = 2**4
args.target_step = args.env.max_step
args.worker_num = 4
train_and_evaluate_mg(args)
if_batch_env = 0
if if_batch_env:
from envs.FinRL.StockTrading import StockVecEnvNAS89
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = int(sys.argv[-1][-4])
args.random_seed += args.gpu_id
args.target_step = args.env.max_step
args.worker_num = 4
train_and_evaluate_mp(args)