def demo4_bullet_mujoco_on_policy():
args = Arguments(if_on_policy=True) # hyper-parameters of on-policy is different from off-policy
import pybullet_envs # for python-bullet-gym
dir(pybullet_envs)
"TotalStep: 1e5, TargetReturn: 18, UsedTime: 3ks, ReacherBulletEnv-v0, PPO"
"TotalStep: 1e6, TargetReturn: 18, UsedTime: 30ks, ReacherBulletEnv-v0, PPO"
args.env = PreprocessEnv(gym.make('ReacherBulletEnv-v0'))
from elegantrl.agent import AgentPPO
args.agent = AgentPPO()
args.agent.if_use_gae = True
args.break_step = int(2e5 * 8)
args.reward_scale = 2 ** 0 # RewardRange: -15 < 0 < 18 < 25
args.gamma = 0.96
args.eval_times1 = 2 ** 2
args.eval_times1 = 2 ** 5
# train_and_evaluate(args)
args.rollout_num = 4
train_and_evaluate_mp(args)
"TotalStep: 3e6, TargetReturn: 1500, UsedTime: 2ks, AntBulletEnv-v0, PPO"
"TotalStep: 10e6, TargetReturn: 2500, UsedTime: 6ks, AntBulletEnv-v0, PPO"
"TotalStep: 46e6, TargetReturn: 3017, UsedTime: 25ks, AntBulletEnv-v0, PPO"
"TotalStep: 5e6, TargetReturn: 1500, UsedTime: 3ks, AntBulletEnv-v0, PPO if_use_dn"
"TotalStep: 15e6, TargetReturn: 2500, UsedTime: 10ks, AntBulletEnv-v0, PPO if_use_dn"
"TotalStep: 60e6, TargetReturn: 2949, UsedTime: 34ks, AntBulletEnv-v0, PPO if_use_dn"
"TotalStep: 2e6, TargetReturn: 1500, UsedTime: 2ks, AntBulletEnv-v0, PPO if_use_cn"
"TotalStep: 10e6, TargetReturn: 2500, UsedTime: 7ks, AntBulletEnv-v0, PPO if_use_cn"
"TotalStep: 53e6, TargetReturn: 2834, UsedTime: 35ks, AntBulletEnv-v0, PPO if_use_cn"
args.env = PreprocessEnv(env=gym.make('AntBulletEnv-v0'))
from elegantrl.agent import AgentPPO
args.agent = AgentPPO()
args.agent.if_use_gae = True
args.agent.lambda_entropy = 0.05
args.agent.lambda_gae_adv = 0.97
args.if_allow_break = False
args.break_step = int(8e6 * 8) # (5e5) 1e6, UsedTime: (15,000s) 30,000s
args.reward_scale = 2 ** -2 # (-50) 0 ~ 2500 (3340)
args.max_memo = args.env.max_step * 4
args.batch_size = 2 ** 11 # 10
args.repeat_times = 2 ** 3
args.eval_gap = 2 ** 8 # for Recorder
args.eva_size1 = 2 ** 1 # for Recorder
args.eva_size2 = 2 ** 3 # for Recorder
# train_and_evaluate(args)
args.rollout_num = 4
train_and_evaluate_mp(args)
def demo2_continuous_action_space_on_policy():
args = Arguments(if_on_policy=True) # hyper-parameters of on-policy is different from off-policy
args.random_seed = 1943
'''choose an DRL algorithm'''
from elegantrl.agent import AgentPPO
args.agent = AgentPPO()
args.agent.if_use_gae = False
'''choose environment'''
"TotalStep: 2e5, TargetReturn: -200, UsedTime: 300s, Pendulum-v0, PPO"
env = gym.make('Pendulum-v0')
env.target_return = -200 # set target_return manually for env 'Pendulum-v0'
args.env = PreprocessEnv(env=env)
args.reward_scale = 2 ** -3 # RewardRange: -1800 < -200 < -50 < 0
args.repeat_times = 2 ** 3
args.target_step = 200 * 8
args.eval_gap = 2 ** 6
"PPO TotalStep: 8e5, TargetReturn: 200, UsedTime: 1500s, LunarLanderContinuous-v2"
# args.env = PreprocessEnv(env=gym.make('LunarLanderContinuous-v2'))
# args.reward_scale = 2 ** 0 # RewardRange: -800 < -200 < 200 < 302
"PPO TotalStep: 8e5, TargetReturn: 300, UsedTime: 1800s, BipedalWalker-v3"
# args.env = PreprocessEnv(env=gym.make('BipedalWalker-v3'))
# args.reward_scale = 2 ** 0 # RewardRange: -200 < -150 < 300 < 334
# args.gamma = 0.96
'''train and evaluate'''
# train_and_evaluate(args)
args.rollout_num = 2
train_and_evaluate_mp(args)
def demo2_continuous_action_space_off_policy():
args = Arguments(if_on_policy=False)
'''choose an DRL algorithm'''
from elegantrl.agent import AgentModSAC # AgentSAC, AgentTD3, AgentDDPG
args.agent = AgentModSAC()
'''choose environment'''
"TotalStep: 3e4, TargetReturn: -200, UsedTime: 300s, Pendulum-v0, TD3"
"TotalStep: 2e4, TargetReturn: -200, UsedTime: 200s, Pendulum-v0, ModSAC "
env = gym.make('Pendulum-v0')
env.target_return = -200 # set target_return manually for env 'Pendulum-v0'
args.env = PreprocessEnv(env=env)
args.reward_scale = 2 ** -3 # RewardRange: -1800 < -200 < -50 < 0
"TD3 TotalStep: 9e4, TargetReturn: 100, UsedTime: 3ks, LunarLanderContinuous-v2"
"TD3 TotalStep: 20e4, TargetReturn: 200, UsedTime: 5ks, LunarLanderContinuous-v2"
"SAC TotalStep: 9e4, TargetReturn: 200, UsedTime: 3ks, LunarLanderContinuous-v2"
"ModSAC TotalStep: 5e4, TargetReturn: 200, UsedTime: 1ks, LunarLanderContinuous-v2"
# args.env = PreprocessEnv(env=gym.make('LunarLanderContinuous-v2'))
# args.reward_scale = 2 ** 0 # RewardRange: -800 < -200 < 200 < 302
# args.eval_times2 = 2 ** 4 # set a large eval_times to get a precise learning curve
"ModSAC TotalStep: 2e5, TargetReturn: 300, UsedTime: 5000s, BipedalWalker-v3"
# args.env = PreprocessEnv(env=gym.make('BipedalWalker-v3'))
# args.reward_scale = 2 ** 0 # RewardRange: -200 < -150 < 300 < 334
# args.net_dim = 2 ** 8
# args.break_step = int(2e5)
# args.if_allow_break = True # allow break training when reach goal (early termination)
# args.break_step = int(2e5 * 4) # break training after 'total_step > break_step'
'''train and evaluate'''
train_and_evaluate(args)
def demo3_custom_env_fin_rl():
from elegantrl.agent import AgentPPO
'''choose an DRL algorithm'''
args = Arguments(if_on_policy=True)
args.agent = AgentPPO()
args.agent.if_use_gae = False
"TotalStep: 5e4, TargetReturn: 1.25, UsedTime: 20s, FinanceStock-v2"
"TotalStep: 20e4, TargetReturn: 1.50, UsedTime: 80s, FinanceStock-v2"
from elegantrl.env import FinanceStockEnv # a standard env for ElegantRL, not need PreprocessEnv()
args.env = FinanceStockEnv(if_train=True, train_beg=0, train_len=1024)
args.env_eval = FinanceStockEnv(if_train=False, train_beg=0, train_len=1024) # eva_len = 1699 - train_len
args.reward_scale = 2 ** 0 # RewardRange: 0 < 1.0 < 1.25 < 1.5 < 1.6
args.break_step = int(5e6)
args.net_dim = 2 ** 8
args.max_step = args.env.max_step
args.max_memo = (args.max_step - 1) * 8
args.batch_size = 2 ** 11
args.repeat_times = 2 ** 4
args.eval_times1 = 2 ** 2
args.eval_times2 = 2 ** 4
args.if_allow_break = True
'''train and evaluate'''
# train_and_evaluate(args)
args.rollout_num = 8
train_and_evaluate_mp(args)
def demo4_bullet_mujoco_off_policy():
args = Arguments(if_on_policy=False)
args.random_seed = 10086
from elegantrl.agent import AgentModSAC # AgentSAC, AgentTD3, AgentDDPG
args.agent = AgentModSAC() # AgentSAC(), AgentTD3(), AgentDDPG()
args.agent.if_use_dn = True
import pybullet_envs # for python-bullet-gym
dir(pybullet_envs)
"TotalStep: 5e4, TargetReturn: 18, UsedTime: 1100s, ReacherBulletEnv-v0"
"TotalStep: 30e4, TargetReturn: 25, UsedTime: s, ReacherBulletEnv-v0"
args.env = PreprocessEnv(gym.make('ReacherBulletEnv-v0'))
args.env.max_step = 2 ** 10 # important, default env.max_step=150
args.reward_scale = 2 ** 0 # -80 < -30 < 18 < 28
args.gamma = 0.96
args.break_step = int(6e4 * 8) # (4e4) 8e5, UsedTime: (300s) 700s
args.eval_times1 = 2 ** 2
args.eval_times1 = 2 ** 5
args.if_per = True
train_and_evaluate(args)
"TotalStep: 3e5, TargetReward: 1500, UsedTime: 4ks, AntBulletEnv-v0 ModSAC if_use_dn"
"TotalStep: 4e5, TargetReward: 2500, UsedTime: 6ks, AntBulletEnv-v0 ModSAC if_use_dn"
"TotalStep: 10e5, TargetReward: 2879, UsedTime: ks, AntBulletEnv-v0 ModSAC if_use_dn"
"TotalStep: 3e5, TargetReward: 1500, UsedTime: 8ks, AntBulletEnv-v0 ModSAC if_use_cn"
"TotalStep: 7e5, TargetReward: 2500, UsedTime: 18ks, AntBulletEnv-v0 ModSAC if_use_cn"
"TotalStep: 16e5, TargetReward: 2923, UsedTime: ks, AntBulletEnv-v0 ModSAC if_use_cn"
args.env = PreprocessEnv(env=gym.make('AntBulletEnv-v0'))
args.break_step = int(6e5 * 8) # (5e5) 1e6, UsedTime: (15,000s) 30,000s
args.if_allow_break = False
args.reward_scale = 2 ** -2 # RewardRange: -50 < 0 < 2500 < 3340
args.max_memo = 2 ** 21
args.batch_size = 2 ** 8
args.repeat_times = 2 ** 1
args.eval_gap = 2 ** 9 # for Recorder
args.eva_size1 = 2 ** 1 # for Recorder
args.eva_size2 = 2 ** 3 # for Recorder
# train_and_evaluate(args)
args.rollout_num = 4
train_and_evaluate_mp(args)
def demo3_custom_env_fin_rl():
from elegantrl.agent import AgentPPO
'''choose an DRL algorithm'''
args = Arguments(if_on_policy=True)
args.agent = AgentPPO()
args.agent.if_use_gae = False
"TotalStep: 5e4, TargetReturn: 1.25, UsedTime: 20s, FinanceStock-v2"
"TotalStep: 20e4, TargetReturn: 1.50, UsedTime: 80s, FinanceStock-v2"
# from elegantrl.env import FinanceStockEnv # a standard env for ElegantRL, not need PreprocessEnv()
# args.env = FinanceStockEnv(if_train=True, train_beg=0, train_len=1024)
# args.env_eval = FinanceStockEnv(if_train=False, train_beg=0, train_len=1024) # eva_len = 1699 - train_len
from finrl.config import config
from beta3 import StockTradingEnv, load_stock_trading_data
train_df, eval_df = load_stock_trading_data()
# train = data_split(processed_df, config.START_DATE, config.START_TRADE_DATE)
# trade = data_split(processed_df, config.START_TRADE_DATE, config.END_DATE)
# calculate state action space
stock_dimension = len(train_df.tic.unique())
state_space = 1 + (2 +
len(config.TECHNICAL_INDICATORS_LIST)) * stock_dimension
env_kwargs = {
"max_stock": 100,
"initial_amount": 1000000,
"buy_cost_pct": 0.001,
"sell_cost_pct": 0.001,
"state_space": state_space,
"stock_dim": stock_dimension,
"tech_indicator_list": config.TECHNICAL_INDICATORS_LIST,
"action_space": stock_dimension,
"reward_scaling": 2**-14
}
args.env = StockTradingEnv(df=train_df, **env_kwargs)
args.env_eval = StockTradingEnv(df=eval_df, **env_kwargs)
args.reward_scale = 2**0 # RewardRange: 0 < 1.0 < 1.25 < 1.5 < 1.6
args.break_step = int(5e6)
args.net_dim = 2**8
args.max_step = args.env.max_step
args.max_memo = (args.max_step - 1) * 8
args.batch_size = 2**11
args.repeat_times = 2**4
args.eval_times1 = 2**1
args.eval_times2 = 2**3
args.if_allow_break = True
'''train and evaluate'''
# train_and_evaluate(args)
args.rollout_num = 4
train_and_evaluate_mp(args)