def demo3_custom_env_fin_rl():
import elegantrl.agent as agent
"""DEMO 3: Custom Continuous action env: FinanceStock-v1"""
args = Arguments(if_on_policy=True)
args.agent = agent.AgentGaePPO() # PPO+GAE (on-policy)
from elegantrl.env import FinanceMultiStockEnv # a standard env for ElegantRL, not need PreprocessEnv()
args.env = FinanceMultiStockEnv(if_train=True, train_beg=0, train_len=1024)
args.env_eval = FinanceMultiStockEnv(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 <
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_break_early = False
"TotalStep: 2e5, TargetReward: 1.25, UsedTime: 200s"
"TotalStep: 4e5, TargetReward: 1.50, UsedTime: 400s"
# train_and_evaluate(args)
args.rollout_num = 8
train_and_evaluate__multiprocessing(args)
def demo3_custom_env_fin_rl():
from elegantrl.agent import AgentGaePPO
'''choose an DRL algorithm'''
args = Arguments(if_on_policy=True)
args.agent = AgentGaePPO() # PPO+GAE (on-policy)
from elegantrl.env import FinanceMultiStockEnv # a standard env for ElegantRL, not need PreprocessEnv()
args.env = FinanceMultiStockEnv(if_train=True, train_beg=0, train_len=1024)
args.env_eval = FinanceMultiStockEnv(
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 <
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 = False
"TotalStep: 2e5, TargetReward: 1.25, UsedTime: 200s"
"TotalStep: 4e5, TargetReward: 1.50, UsedTime: 400s"
"TotalStep: 10e5, TargetReward: 1.62, UsedTime: 1000s"
'''train and evaluate'''
# train_and_evaluate(args)
args.rollout_num = 8
train_and_evaluate__multiprocessing(args)
def run__demo():
import gym
gym.logger.set_level(40) # Block warning: 'WARN: Box bound precision lowered by casting to float32'
from elegantrl.env import decorate_env, FinanceMultiStockEnv
args = Arguments(agent_rl=None, env=None, gpu_id=None) # see Arguments() to see hyper-parameters
'''DEMO 1: Discrete action env: CartPole-v0 of gym'''
from elegantrl.agent import AgentDoubleDQN
args.agent_rl = AgentDoubleDQN # choose an DRL algorithm
args.env = decorate_env(env=gym.make('CartPole-v0'))
args.net_dim = 2 ** 7 # change a default hyper-parameters
args.env = decorate_env(env=gym.make('LunarLander-v2'))
args.net_dim = 2 ** 8 # change a default hyper-parameters
train_and_evaluate(args)
exit()
'''DEMO 2: Continuous action env: gym.box2D'''
from elegantrl.agent import AgentSAC # AgentTD3
args.agent_rl = AgentSAC # off-policy
env = gym.make('Pendulum-v0')
env.target_reward = -200 # set target_reward manually for env 'Pendulum-v0'
args.env = decorate_env(env=env)
# args.env = decorate_env(env=gym.make('LunarLanderContinuous-v2'))
# args.env = decorate_env(env=gym.make('BipedalWalker-v3')) # recommend args.gamma = 0.95
train_and_evaluate(args)
exit()
from elegantrl.agent import AgentGaePPO # AgentPPO
args.agent_rl = AgentGaePPO # on-policy
args = Arguments(if_on_policy=True) # on-policy has different hyper-parameters from off-policy
env = gym.make('Pendulum-v0')
env.target_reward = -200 # set target_reward manually for env 'Pendulum-v0'
args.env = decorate_env(env=env)
# args.env = decorate_env(env=gym.make('LunarLanderContinuous-v2'))
# args.env = decorate_env(env=gym.make('BipedalWalker-v3')) # recommend args.gamma = 0.95
train_and_evaluate(args)
exit()
'''DEMO 3: Custom Continuous action env: FinanceStock-v1'''
args = Arguments(if_on_policy=True)
from elegantrl.agent import AgentGaePPO # AgentPPO
args.agent_rl = AgentGaePPO # PPO+GAE (on-policy)
from elegantrl.env import FinanceMultiStockEnv
args.env = FinanceMultiStockEnv() # a standard env for ElegantRL, not need decorate_env()
args.break_step = int(5e6 * 4) # 5e6 (15e6) UsedTime 3,000s (9,000s)
args.net_dim = 2 ** 8
args.max_step = 1699
args.max_memo = (args.max_step - 1) * 16
args.batch_size = 2 ** 11
args.repeat_times = 2 ** 4
train_and_evaluate(args)
exit()
def run__demo():
import elegantrl.BetaWarning.agent as agent
from elegantrl.BetaWarning.env import prep_env
# from elegantrl.main import Arguments, train_and_evaluate, train_and_evaluate__multiprocessing
import gym
gym.logger.set_level(
40
) # Block warning: 'WARN: Box bound precision lowered by casting to float32'
"""DEMO 1: Discrete action env: CartPole-v0 of gym"""
args = Arguments(agent_rl=None, env=None,
gpu_id=None) # see Arguments() to see hyper-parameters
args.agent_rl = agent.AgentD3QN # choose an DRL algorithm
args.env = prep_env(env=gym.make('CartPole-v0'))
args.net_dim = 2**7 # change a default hyper-parameters
# args.env = decorate_env(env=gym.make('LunarLander-v2'))
# args.net_dim = 2 ** 8 # change a default hyper-parameters
train_and_evaluate(args)
"""DEMO 2: Continuous action env, gym.Box2D"""
if_on_policy = False
args = Arguments(
if_on_policy=if_on_policy
) # on-policy has different hyper-parameters from off-policy
if if_on_policy:
args.agent_rl = agent.AgentGaePPO # on-policy: AgentPPO, AgentGaePPO
else:
args.agent_rl = agent.AgentModSAC # off-policy: AgentSAC, AgentModPPO, AgentTD3, AgentDDPG
env = gym.make('Pendulum-v0')
env.target_reward = -200 # set target_reward manually for env 'Pendulum-v0'
args.env = prep_env(env=env)
args.net_dim = 2**7 # change a default hyper-parameters
# args.env = decorate_env(env=gym.make('LunarLanderContinuous-v2'))
# args.env = decorate_env(env=gym.make('BipedalWalker-v3')) # recommend args.gamma = 0.95
train_and_evaluate(args)
"""DEMO 3: Custom Continuous action env: FinanceStock-v1"""
args = Arguments(if_on_policy=True)
args.agent_rl = agent.AgentGaePPO # PPO+GAE (on-policy)
from elegantrl.env import FinanceMultiStockEnv
args.env = FinanceMultiStockEnv(
if_train=True) # a standard env for ElegantRL, not need decorate_env()
args.env_eval = FinanceMultiStockEnv(if_train=False)
args.break_step = int(5e6) # 5e6 (15e6) UsedTime 3,000s (9,000s)
args.net_dim = 2**8
args.target_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**4
# train_and_evaluate(args)
args.rollout_num = 4
train_and_evaluate__multiprocessing(args)
args.env_eval.draw_cumulative_return(args, torch)
'''DEMO 4: PyBullet(MuJoCo) Robot Env'''
args = Arguments(if_on_policy=True)
args.agent_rl = agent.AgentGaePPO # agent.AgentPPO
import pybullet_envs # for python-bullet-gym
dir(pybullet_envs)
# args.env = decorate_env(gym.make('AntBulletEnv-v0'))
args.env = prep_env(gym.make('ReacherBulletEnv-v0'))
# args.repeat_times=8
# args.max_memo=args.target_step =4096
args.break_step = int(4e8) # (4e4) 8e5, UsedTime: (300s) 700s
args.if_break_early = False
args.eval_times1 = 2**2
args.eval_times1 = 2**4
args.rollout_num = 4
train_and_evaluate__multiprocessing(args)
# """DEMO 5: Discrete action env: CartPole-v0 of gym"""
import pybullet_envs # for python-bullet-gym
args = Arguments(agent_rl=None, env=None,
gpu_id=0) # see Arguments() to see hyper-parameters
args.agent_rl = agent.AgentTD3 # choose an DRL algorithm
args.env = prep_env(env=gym.make('ReacherBulletEnv-v0'))
args.net_dim = 2**7 # change a default hyper-parameters
args.if_per = True
args.break_step = int(2e20) # (4e4) 8e5, UsedTime: (300s) 700s
# train_and_evaluate(args)
train_and_evaluate__multiprocessing(args)
exit(0)
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: 'WARN: Box bound precision lowered by casting to float32'
"""DEMO 1: Discrete action env of gym"""
args = Arguments(agent=None, env=None,
gpu_id=None) # see Arguments() to see hyper-parameters
'''choose an DRL algorithm'''
from elegantrl.tutorial.agent import AgentDoubleDQN # AgentDQN
args.agent = AgentDoubleDQN()
'''choose environment'''
# args.env = PreprocessEnv(env=gym.make('CartPole-v0'))
# args.net_dim = 2 ** 7 # change a default hyper-parameters
# args.batch_size = 2 ** 7
"TotalStep: 2e3, TargetReward: , UsedTime: 10s"
args.env = PreprocessEnv(env=gym.make('LunarLander-v2'))
args.net_dim = 2**8
args.batch_size = 2**8
"TotalStep: 6e4, TargetReward: 200, UsedTime: 600s"
'''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()
'''choose environment'''
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.net_dim = 2**7
args.batch_size = 2**7
"TotalStep: 4e5, TargetReward: -200, UsedTime: 400s"
# args.env = PreprocessEnv(env=gym.make('LunarLanderContinuous-v2'))
# args.reward_scale = 2 ** 0 # RewardRange: -800 < -200 < 200 < 302
"TotalStep: 9e4, TargetReward: 200, UsedTime: 2500s"
# args.env = PreprocessEnv(env=gym.make('BipedalWalker-v3'))
# args.reward_scale = 2 ** 0 # RewardRange: -200 < -150 < 300 < 334
# args.break_step = int(2e5)
# args.if_allow_break = False
"TotalStep: 2e5, TargetReward: 300, UsedTime: 5000s"
'''train and evaluate'''
train_and_evaluate(args)
# train_and_evaluate__multiprocessing(args) # try multiprocessing in complete version
exit()
'''DEMO 3: Custom Continuous action env: FinanceStock-v1'''
args = Arguments(if_on_policy=True)
'''choose an DRL algorithm'''
from elegantrl.tutorial.agent import AgentPPO
args.agent = AgentPPO()
from elegantrl.env import FinanceMultiStockEnv # a standard env for ElegantRL, not need PreprocessEnv()
args.env = FinanceMultiStockEnv(if_train=True, train_beg=0, train_len=1024)
args.env_eval = FinanceMultiStockEnv(
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 <
args.break_step = int(5e6)
args.max_step = args.env.max_step
args.max_memo = (args.max_step - 1) * 8
args.batch_size = 2**11
args.if_allow_break = False
"TotalStep: 2e5, TargetReward: 1.25, UsedTime: 200s"
"TotalStep: 4e5, TargetReward: 1.50, UsedTime: 400s"
"TotalStep: 10e5, TargetReward: 1.62, UsedTime: 1000s"
'''train and evaluate'''
train_and_evaluate(args)
# args.rollout_num = 8
# train_and_evaluate__multiprocessing(args) # try multiprocessing in complete version
exit()