Pendulum ~~~~~~~~~~~~~~~~~~ Overview ========== The inverted pendulum problem is a classic control problem in reinforcement learning. Pendulum is a continuous control task in the inverted pendulum problem. The pendulum starts at a random position and the goal is to swing up to stay upright. As shown below. .. image:: ./images/pendulum.gif :align: center Install ======== Installation Method -------------------- The Pendulum environment is built into the gym, and you can install the gym directly. Its environment id is \ ``Pendulum-v1`` \. .. code:: shell pip install gym Verify Installation -------------------- Run the following Python program, if no error is reported, the installation is successful. .. code:: shell import gym env = gym.make('Pendulum-v1') obs = env.reset() print(obs) Environment Introduction ========================= Action Space --------------- The action space of Pendulum belongs to the continuous action space. - \ ``Control torque`` \: The size range is \ ``[-2, 2]`` \. Using the gym environment space definition can be expressed as: .. code:: python action_space = spaces.Box(low=-2,high=2) State Space ------------ The state space of Pendulum has 3 elements that describe the angle and angular velocity of the pendulum. They are: - \ ``sin`` \: The sin value of the angle the pendulum deviates from the vertical direction, the range is \ ``[-1, 1]`` \. - \ ``cos`` \: The cos value of the angle the pendulum deviates from the vertical direction, the range is \ ``[-1, 1]`` \. - \ ``thetadot`` \: Angular angle of the pendulum, in the range \ ``[-8, 8]`` \. Reward Space ------------- First calculate \ ``cost`` \, including three terms: - \ ``angle_normalize(th)**2`` \: Penalty for the angle difference between the current pendulum and the target position - \ ``0.1*thdot**2`` \: Penalty for angular velocity. Avoid approaching the target while still having a large angular velocity, thus overshooting the target position. - \ ``0.001*(u**2)`` \: Penalty for input torque. The bigger the moment we use, the bigger the penalty. Add the three terms to get \ ``cost`` \. Finally, the inverse of \ ``cost`` \, which is \ ``-cost`` \, is returned as the reward value. Termination Condition ---------------------- The termination condition for each episode of the Pendulum environment is any of the following: - Reach the maximum step of the episode. Other ======= Store Video ------------ Some environments have their own rendering plugins. DI-engine does not support the rendering plug-in that comes with the environment, but generates video recordings by saving the logs during training. For details, please refer to the Visualization & Logging section under the DI-engine `official documentation `__ Quick start chapter. DI-zoo Runnable Code Example ============================= The following provides a complete Pendulum environment config, using the DDPG algorithm as the policy. Please run the \ ``pendulum_ddpg_main.py`` \ file in the \ ``DI-engine/dizoo/classic_control/pendulum/entry`` \ directory, as follows. .. code:: python import os import gym from tensorboardX import SummaryWriter from ding.config import compile_config from ding.worker import BaseLearner, SampleSerialCollector, InteractionSerialEvaluator, AdvancedReplayBuffer from ding.envs import BaseEnvManager, DingEnvWrapper from ding.policy import DDPGPolicy from ding.model import QAC from ding.utils import set_pkg_seed from dizoo.classic_control.pendulum.envs import PendulumEnv from dizoo.classic_control.pendulum.config.pendulum_ddpg_config import pendulum_ddpg_config def main(cfg, seed=0): cfg = compile_config( cfg, BaseEnvManager, DDPGPolicy, BaseLearner, SampleSerialCollector, InteractionSerialEvaluator, AdvancedReplayBuffer, save_cfg=True ) # Set up envs for collection and evaluation collector_env_num, evaluator_env_num = cfg.env.collector_env_num, cfg.env.evaluator_env_num collector_env = BaseEnvManager( env_fn=[lambda: PendulumEnv(cfg.env) for _ in range(collector_env_num)], cfg=cfg.env.manager ) evaluator_env = BaseEnvManager( env_fn=[lambda: PendulumEnv(cfg.env) for _ in range(evaluator_env_num)], cfg=cfg.env.manager ) # Set random seed for all package and instance collector_env.seed(seed) evaluator_env.seed(seed, dynamic_seed=False) set_pkg_seed(seed, use_cuda=cfg.policy.cuda) # Set up RL Policy model = QAC(**cfg.policy.model) policy = DDPGPolicy(cfg.policy, model=model) # Set up collection, training and evaluation utilities tb_logger = SummaryWriter(os.path.join('./{}/log/'.format(cfg.exp_name), 'serial')) learner = BaseLearner(cfg.policy.learn.learner, policy.learn_mode, tb_logger, exp_name=cfg.exp_name) collector = SampleSerialCollector( cfg.policy.collect.collector, collector_env, policy.collect_mode, tb_logger, exp_name=cfg.exp_name ) evaluator = InteractionSerialEvaluator( cfg.policy.eval.evaluator, evaluator_env, policy.eval_mode, tb_logger, exp_name=cfg.exp_name ) replay_buffer = AdvancedReplayBuffer(cfg.policy.other.replay_buffer, tb_logger, exp_name=cfg.exp_name) # Training & Evaluation loop while True: # Evaluate at the beginning and with specific frequency if evaluator.should_eval(learner.train_iter): stop, reward = evaluator.eval(learner.save_checkpoint, learner.train_iter, collector.envstep) if stop: break # Collect data from environments new_data = collector.collect(train_iter=learner.train_iter) replay_buffer.push(new_data, cur_collector_envstep=collector.envstep) # Train for i in range(cfg.policy.learn.update_per_collect): train_data = replay_buffer.sample(learner.policy.get_attribute('batch_size'), learner.train_iter) if train_data is None: break learner.train(train_data, collector.envstep) if __name__ == "__main__": main(pendulum_ddpg_config, seed=0) Experimental Results ===================== The experimental results using the DDPG algorithm are as follows. The abscissa is \ ``episode`` \, and the ordinate is \ ``reward_mean`` \. .. image:: ./images/pendulum_ddpg.png :align: center :scale: 20 % References ====================== - Pendulum `source code `__