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, TargetReward: 1.25, UsedTime: 20s, FinanceStock-v2"
"TotalStep: 20e4, TargetReward: 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 run__demo():
# from elegantrl.tutorial.run import Arguments, train_and_evaluate
from elegantrl.tutorial.env import PreprocessEnv
import gym
gym.logger.set_level(40) # Block warning
"""DEMO 1: Discrete action env of gym"""
args = Arguments(agent=None, env=None, gpu_id=None) # hyperparameters
'''choose an DRL algorithm'''
from elegantrl.tutorial.agent import AgentDoubleDQN # AgentDQN
args.agent = AgentDoubleDQN()
'''choose environment'''
# "TotalStep: 2e3, TargetReward: 195, UsedTime: 10s, CartPole-v0"
# args.env = PreprocessEnv(env=gym.make('CartPole-v0'))
# args.net_dim = 2 ** 7 # change a default hyper-parameters
# args.batch_size = 2 ** 7
"TotalStep: 6e4, TargetReward: 200, UsedTime: 600s, LunarLander-v2"
args.env = PreprocessEnv(env=gym.make('LunarLander-v2'))
args.net_dim = 2**8
args.batch_size = 2**8
'''train and evaluate'''
train_and_evaluate(args)
exit()
'''DEMO 2: Continuous action env if gym'''
'''DEMO 2.1: choose an off-policy DRL algorithm'''
from elegantrl.tutorial.agent import AgentSAC # AgentTD3, AgentDDPG
args = Arguments(if_on_policy=False)
args.agent = AgentSAC()
'''DEMO 2.2: choose an on-policy DRL algorithm'''
from elegantrl.tutorial.agent import AgentPPO # AgentGaePPO
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
args.agent = AgentPPO()
"TotalStep: 4e5, TargetReward: -200, UsedTime: 400s, Pendulum-v0"
env = gym.make('Pendulum-v0')
env.target_reward = -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.batch_size = 2**7
args.net_dim = 2**7
"TotalStep: 9e4, TargetReward: 200, UsedTime: 2500s, LunarLanderContinuous-v2"
# args.env = PreprocessEnv(env=gym.make('LunarLanderContinuous-v2'))
# args.reward_scale = 2 ** 0 # RewardRange: -800 < -200 < 200 < 302
"TotalStep: 2e5, TargetReward: 300, UsedTime: 5000s, BipedalWalker-v3"
# args.env = PreprocessEnv(env=gym.make('BipedalWalker-v3'))
# args.reward_scale = 2 ** 0 # RewardRange: -200 < -150 < 300 < 334
# args.break_step = int(2e5) # break training when reach break_step (force termination)
# args.if_allow_break = False # allow break training when reach goal (early termination)
train_and_evaluate(args) # tutorial version
exit()
'''DEMO 3: Custom Continuous action env: FinanceStock-v1'''
from elegantrl.tutorial.agent import AgentPPO
args = Arguments(if_on_policy=True)
args.agent = AgentPPO()
args.agent.if_use_gae = False
"TotalStep: 5e4, TargetReward: 1.25, UsedTime: 30s, FinanceStock-v2"
"TotalStep: 16e4, TargetReward: 1.50, UsedTime: 160s, 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.env_eval.target_reward = 1.25 # denotes 1.25 times the initial_account. convergence to 1.5
args.break_step = int(
5e6) # break training when reach break_step (force termination)
args.if_allow_break = True # allow break training when reach goal (early termination)
args.batch_size = 2**11
args.max_step = args.env.max_step
args.max_memo = (args.max_step - 1) * 8
train_and_evaluate(args) # tutorial version
# train_and_evaluate_mp(args) # try multiprocessing in advanced version
exit()