def demo1_discrete_action_space():
args = Arguments(agent=None, env=None, gpu_id=None) # see Arguments() to see hyper-parameters
'''choose an DRL algorithm'''
# from elegantrl.agent import AgentD3QN # AgentDQN,AgentDuelDQN, AgentDoubleDQN,
# args.agent = AgentD3QN()
from elegantrl.agent import AgentDuelingDQN # AgentDQN,AgentDuelDQN, AgentDoubleDQN,
args.agent = AgentDuelingDQN()
'''choose environment'''
"TotalStep: 2e3, TargetReturn: 200, UsedTime: 20s, CartPole-v0"
"TotalStep: 2e3, TargetReturn: 200, UsedTime: 30s, CartPole-v0 rollout_num = 2"
# args.env = PreprocessEnv(env=gym.make('CartPole-v0'))
# args.net_dim = 2 ** 7 # change a default hyper-parameters
# args.batch_size = 2 ** 7
# args.target_step = 2 ** 8
# args.eval_gap = 2 ** 0
"TotalStep: 6e4, TargetReturn: 200, UsedTime: 600s, LunarLander-v2, D3DQN"
"TotalStep: 4e4, TargetReturn: 200, UsedTime: 600s, LunarLander-v2, DuelDQN"
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)
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 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)
if __name__ == '__main__':
args = Arguments(if_on_policy=True)
args.agent = AgentPPO()
args.agent.if_use_gae = True
args.agent.lambda_entropy = 0.04
from kuka_cam_reach_env import KukaCamReachEnv, CustomSkipFrame
env_config = {
"is_render": False,
"is_good_view": False,
"max_steps_one_episode": 1000,
}
args.env = CustomSkipFrame(KukaCamReachEnv(config=env_config))
args.gamma = 0.995
args.break_step = int(3e5)
args.net_dim = 2**9
args.max_step = args.env.max_step
args.max_memo = args.max_step * 4
args.batch_size = 2**10
args.repeat_times = 2**3
args.eval_gap = 2**4
args.eval_times1 = 2**3
args.eval_times2 = 2**5
args.if_allow_break = False
'''train and evaluate'''
# train_and_evaluate(args)
args.rollout_num = 1
train_and_evaluate(args)
def train_model(self, model, cwd, total_timesteps=5000):
model.cwd = cwd
model.break_step = total_timesteps
train_and_evaluate(model)