Continuous action example for IMPALA

[supersglzc]

• I have marked all applicable categories:
  • exception-raising bug
  • RL algorithm bug
  • system worker bug
  • system utils bug
  • code design/refactor
  • documentation request
  • new feature request
• I have visited the readme and doc
• I have searched through the issue tracker and pr tracker
• I have mentioned version numbers, operating system and environment, where applicable:

  import ding, torch, sys
  print(ding.__version__, torch.__version__, sys.version, sys.platform)

Hi, I find that DI-Engine aims to support IMPALA for both continuous and discrete action spaces, however, there is no config examples for continuous action tasks. One thing I notice is that IMPALA uses arg-max for evaluation, which I suppose to be designed for discrete action tasks specifically? I am wondering if we can have some demos of IMPALA on continuous action tasks?

I tried to make one, but it fails to run:

from easydict import EasyDict

bipedalwalker_impala_config = dict(
    exp_name='bipedalwalker_impala_seed0',
    env=dict(
        env_id='BipedalWalker-v3',
        collector_env_num=8,
        evaluator_env_num=5,
        # (bool) Scale output action into legal range.
        act_scale=True,
        n_evaluator_episode=5,
        stop_value=300,
        rew_clip=True,
        # The path to save the game replay
        # replay_path='./bipedalwalker_ppo_seed0/video',
    ),
    policy=dict(
        cuda=False,
        action_space='continuous',
        model=dict(
            action_space='continuous',
            obs_shape=24,
            action_shape=4,
        ),
        learn=dict(
            # (int) collect n_sample data, train model update_per_collect times
            # here we follow ppo serial pipeline
            update_per_collect=4,
            # (int) the number of data for a train iteration
            batch_size=16,
            learning_rate=0.0005,
            # (float) loss weight of the value network, the weight of policy network is set to 1
            value_weight=0.5,
            # (float) loss weight of the entropy regularization, the weight of policy network is set to 1
            entropy_weight=0.0001,
            # (float) discount factor for future reward, defaults int [0, 1]
            discount_factor=0.9,
            # (float) additional discounting parameter
            lambda_=0.95,
            # (int) the trajectory length to calculate v-trace target
            unroll_len=32,
            # (float) clip ratio of importance weights
            rho_clip_ratio=1.0,
            # (float) clip ratio of importance weights
            c_clip_ratio=1.0,
            # (float) clip ratio of importance sampling
            rho_pg_clip_ratio=1.0,
        ),
        collect=dict(
            # (int) collect n_sample data, train model n_iteration times
            n_sample=16,
            # (int) the trajectory length to calculate v-trace target
            unroll_len=32,
            # (float) discount factor for future reward, defaults int [0, 1]
            discount_factor=0.9,
            gae_lambda=0.95,
            collector=dict(collect_print_freq=1000, ),
        ),
        eval=dict(evaluator=dict(eval_freq=200, )),
        other=dict(replay_buffer=dict(
            replay_buffer_size=1000,
            max_use=16,
        ), ),
    ),
)
bipedalwalker_impala_config = EasyDict(bipedalwalker_impala_config)
main_config = bipedalwalker_impala_config
bipedalwalker_impala_create_config = dict(
    env=dict(
        type='bipedalwalker',
        import_names=['dizoo.box2d.bipedalwalker.envs.bipedalwalker_env'],
    ),
    env_manager=dict(type='base'),
    policy=dict(type='impala'),
)
bipedalwalker_impala_create_config = EasyDict(bipedalwalker_impala_create_config)
create_config = bipedalwalker_impala_create_config

if __name__ == "__main__":
    # or you can enter `ding -m serial_onpolicy -c bipedalwalker_impala_config.py -s 0`
    from ding.entry import serial_pipeline_onpolicy
    serial_pipeline_onpolicy([main_config, create_config], seed=0)

[PaParaZz1]

The current implementation only supports discrete action space, we will add a continuous version and related examples next week. You can continue to follow this issue.

[PaParaZz1]

This issue has been solved in #551