def demo1__discrete_action_space():
"""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 = 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)
def demo2():
"""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 = decorate_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)
def demo5():
args = Arguments(if_on_policy=False)
# args.agent_rl = agent.AgentModSAC
args.agent_rl = agent.AgentInterSAC
import pybullet_envs # for python-bullet-gym
dir(pybullet_envs)
args.env = decorate_env(gym.make('AntBulletEnv-v0'))
# args.env = decorate_env(gym.make('ReacherBulletEnv-v0'))
args.break_step = int(1e6 * 8) # (5e5) 1e6, UsedTime: (15,000s) 30,000s
args.reward_scale = 2 ** -2 # (-50) 0 ~ 2500 (3340)
args.max_memo = 2 ** 19
args.net_dim = 2 ** 7 # todo
args.eva_size = 2 ** 5 # for Recorder
args.show_gap = 2 ** 8 # for Recorder
train_and_evaluate(args)