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_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 check_stock_trading_env():
if_eval = True # False
env = StockTradingEnv(if_eval=if_eval)
action_dim = env.action_dim
state = env.reset()
print('| check_stock_trading_env, state_dim', len(state))
from time import time
timer = time()
# ============================================================
policy_name = 'Random Action 1e-2'
step = 1
done = False
episode_return = 0
# state = env.reset()
while not done:
action = rd.uniform(-1, 1, size=action_dim) * 1e-2
next_state, reward, done, _ = env.step(action)
# print(';', len(next_state), env.day, reward)
episode_return += reward
step += 1
print()
print(f"| {policy_name}:")
print(f"| step {step}, UsedTime {time() - timer:.3e}")
print(f"| gamma_reward \t\t\t{env.gamma_reward:.3e}")
print(f"| episode return \t\t{episode_return:.3e}")
print(
f"| discount return \t\t{episode_return / step / (1 - env.gamma):.3e}")
print(f"| env episode return \t{env.episode_return:.3e}")
# ============================================================
policy_name = 'Buy 4 Action'
step = 1
done = False
episode_return = 0
# state = env.reset()
while not done:
action = np.zeros(action_dim)
action[:3] = 1
next_state, reward, done, _ = env.step(action)
# print(';', len(next_state), env.day, reward)
episode_return += reward
step += 1
print()
print(f"| {policy_name}:")
print(f"| step {step}, UsedTime {time() - timer:.3e}")
print(f"| gamma_reward \t\t\t{env.gamma_reward:.3e}")
print(f"| episode return \t\t{episode_return:.3e}")
print(
f"| discount return \t\t{episode_return / step / (1 - env.gamma):.3e}")
print(f"| env episode return \t{env.episode_return:.3e}")
# ============================================================
'''draw_cumulative_return'''
from elegantrl2.agent import AgentPPO
from elegantrl2.run import Arguments
args = Arguments(if_on_policy=True)
args.agent = AgentPPO()
args.env = StockTradingEnv(if_eval=True)
args.if_remove = False
args.cwd = './StockTradingEnv-v1_AgentPPO'
args.init_before_training()
env.draw_cumulative_return(args, torch)
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_custom_env_finance_rl_dow30(): # 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.cri_target = True
args.agent.lambda_entropy = 0.02
args.gamma = 0.995
from envs.FinRL.StockTrading import StockEnvDOW30, StockVecEnvDOW30
args.env = StockEnvDOW30(if_eval=False, gamma=args.gamma)
args.env_eval = StockEnvDOW30(if_eval=True, gamma=args.gamma)
args.repeat_times = 2**4
args.learning_rate = 2**-14
args.net_dim = 2**8
args.batch_size = args.net_dim
args.eval_gap = 2**7
args.eval_times1 = 2**0
args.eval_times2 = 2**1
args.break_step = int(10e6)
args.if_allow_break = False
if_single_env = 0
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 * 4
args.worker_num = 4
train_and_evaluate_mp(args)
if_batch_env = 1
if if_batch_env:
args.env = StockVecEnvDOW30(if_eval=False, gamma=args.gamma, env_num=4)
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)
if_multi_learner = 0
if if_multi_learner:
args.env = StockVecEnvDOW30(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = (0, 1)
args.worker_num = 2
train_and_evaluate_mg(args)
def demo_custom_env_finance_rl_nas89():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
args.random_seed = 1943
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.cri_target = True
args.agent.lambda_entropy = 0.04
from envs.FinRL.StockTrading import StockEnvDOW30, StockEnvNAS89, StockVecEnvNAS89
args.gamma = 0.999
if_dow30_daily = 1
if if_dow30_daily:
args.env = StockEnvDOW30(if_eval=False, gamma=args.gamma)
args.env_eval = StockEnvDOW30(if_eval=True, gamma=args.gamma)
else: # elif if_nas89_minute:
args.env = StockEnvNAS89(if_eval=False, gamma=args.gamma)
args.env_eval = StockEnvNAS89(if_eval=True, gamma=args.gamma)
args.repeat_times = 2**4
args.learning_rate = 2**-14
args.net_dim = int(2**8 * 1.5)
args.batch_size = args.net_dim * 4
args.target_step = args.env.max_step
args.eval_gap = 2**8
args.eval_times1 = 2**0
args.eval_times2 = 2**1
args.break_step = int(16e6)
args.if_allow_break = False
if_single_env = 1
if if_single_env:
args.gpu_id = 0
args.worker_num = 4
train_and_evaluate_mp(args)
if_batch_env = 0
if if_batch_env:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = 3
args.random_seed += args.gpu_id
args.worker_num = 2
train_and_evaluate_mp(args)
if_multi_learner = 0
if if_multi_learner:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = (0, 1)
args.worker_num = 2
train_and_evaluate_mg(args)
def demo_custom_env_finance_rl():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
args.random_seed = 0
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.cri_target = True
args.agent.lambda_entropy = 0.04
from envs.FinRL.StockTrading import StockEnvNAS89, StockVecEnvNAS89
args.gamma = 0.999
args.env = StockEnvNAS89(if_eval=False, gamma=args.gamma)
args.env_eval = StockEnvNAS89(if_eval=True, gamma=args.gamma)
args.repeat_times = 2**4
args.learning_rate = 2**-14
args.net_dim = int(2**8 * 1.5)
args.batch_size = args.net_dim * 4
if_single_env = 0
if if_single_env:
args.gpu_id = 0
args.worker_num = 4
train_and_evaluate_mp(args)
if_batch_env = 1
if if_batch_env:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = 0
args.worker_num = 2
train_and_evaluate_mp(args)
if_multi_learner = 0
if if_multi_learner:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = (0, 1)
args.worker_num = 2
train_and_evaluate_mg(args)
"TotalStep: 52e5, TargetReturn: 2.35, UsedTime: 3934s, FinanceStock-v2"
"TotalStep: 81e5, TargetReturn: 2.47, UsedTime: 6129s, FinanceStock-v2"
"TotalStep: 19e5, TargetReturn: 2.50, UsedTime: 1654s, FinanceStock-v2 GPU 2, 3"
"TotalStep: 65e5, TargetReturn: 4.61, UsedTime: 5659s, FinanceStock-v2 GPU 2, 3"
"TotalStep: 18e5, TargetReturn: 2.50, UsedTime: 1452s, FinanceStock-v2 GPU 0, 1"
"TotalStep: 61e5, TargetReturn: 3.92, UsedTime: 4921s, FinanceStock-v2 GPU 0, 1"
"TotalStep: 4e5, TargetReturn: 2.20, UsedTime: 583s, FinanceStock-v2 GPU 0, 1, 2, 3"
"TotalStep: 11e6, TargetReturn: 4.39, UsedTime: 9648s, FinanceStock-v2 GPU 0, 1, 2, 3"
def demo_continuous_action_on_policy():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.gpu_id = sys.argv[-1][-4]
args.agent.cri_target = True
args.learning_rate = 2**-14
args.random_seed = 1943
'''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: 6e5, TargetReward: 200, UsedTime: 800s"
env_name = 'LunarLanderContinuous-v2'
# args.env = PreprocessEnv(env=env_name)
args.env = PreprocessVecEnv(env=env_name, env_num=2)
args.env_eval = PreprocessEnv(env=env_name)
args.reward_scale = 2**0 # RewardRange: -800 < -200 < 200 < 302
args.break_step = int(8e6)
args.if_per_or_gae = True
args.target_step = args.env.max_step * 8
args.repeat_times = 2**4
if_train_bipedal_walker = 1
if if_train_bipedal_walker:
"TotalStep: 8e5, TargetReward: 300, UsedTime: 1800s"
env_name = 'BipedalWalker-v3'
# args.env = PreprocessEnv(env=env_name)
args.env = PreprocessVecEnv(env=env_name, env_num=2)
args.env_eval = PreprocessEnv(env=env_name)
args.reward_scale = 2**0 # RewardRange: -200 < -150 < 300 < 334
args.gamma = 0.97
args.target_step = args.env.max_step * 8
args.repeat_times = 2**4
args.if_per_or_gae = True
args.agent.lambda_entropy = 0.05
args.break_step = int(8e6)
'''train and evaluate'''
# train_and_evaluate(args)
args.worker_num = 2
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)
def demo_custom_env_finance_rl():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
args.random_seed = 0
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.cri_target = True
args.agent.lambda_entropy = 0.04
from envs.FinRL.StockTrading import StockEnvNAS89, StockVecEnvNAS89
args.gamma = 0.999
args.env = StockEnvNAS89(if_eval=False, gamma=args.gamma)
args.eval_env = StockEnvNAS89(if_eval=True, gamma=args.gamma)
args.repeat_times = 2**4
args.learning_rate = 2**-14
args.net_dim = int(2**8 * 1.5)
args.batch_size = args.net_dim * 4
if_single_env = 0
if if_single_env:
args.gpu_id = 0
args.worker_num = 4
train_and_evaluate_mp(args)
if_batch_env = 1
if if_batch_env:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = 0
args.worker_num = 2
train_and_evaluate_mp(args)
if_multi_learner = 0
if if_multi_learner:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = (0, 1)
args.worker_num = 2
train_and_evaluate_mg(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)
