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 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)