kngwyu / rainy Goto Github PK

Now we have two types of agents

• OneStepAgent
  • for DQN-like algorithms
  • execute 1-step + stores transition to replay buffer + train agent by sampled transitions
• NStepParallelAgent
  • for A2C-like algorithms
  • execute N-step in parallel environments + train the policy in an online manner
    These 2 divisions are practical but lack flexibility.
    E.g., we cannot extend OneStep algorithms to batched-parallel style without rewriting the whole process.

So we should re-define agent hierarchies using some important properties, like

• Online/Offline(or use replay buffer or not)
• MultiStep/OneStep
• Not Parallel/Batch Parallel/ Async Parallel

• Use pandas/xarray instead of hand-made Log object
• Integration with Jupyter
• Real time experiment tracking

To correctly evaluate PyBullet envs.

Our CLI tool should be more customizable like RLPy3's one.

There's environments only with VecEnv(like coinrun) so it's desirable to enable to set None as eval_env.

They subtracts (initial - minimum) / max_steps from the value every time they're called, but they are called only max_steps / (nstpes * nworkers) times in A2C, PPO and ACKTR.

This should be fixed and the benchmark in #8 should be done again.

rms.py and atari_wrappers.py are copied from openai baselines, and not well written.
They need refactoring.

Now they're only in parallel wrappers

https://github.com/openai/gym#what-s-new
It is introduced in v0.15.2.
Now we use wrappers ported from baselines, but it's good time to migrate.

Currently loss is reported occasionally but we should compute running stats of loss and report it.

It may be useful for debugging and parameter tuining

Hi @kngwyu,

When I ran your bootdqn_cartpole.py like this:

python examples\bootdqn_cartpole.py train

I got:

Traceback (most recent call last):
File "examples\bootdqn_cartpole.py", line 21, in
) -> Config:
File "C:\rainy\cli.py", line 227, in decorator
rainy_cli(obj=_CLIContext(f, agent, agent_selector, script_path))
File "C:\Miniconda3\lib\site-packages\click\core.py", line 829, in call
return self.main(*args, **kwargs)
File "C:\Miniconda3\lib\site-packages\click\core.py", line 782, in main
rv = self.invoke(ctx)
File "C:\Miniconda3\lib\site-packages\click\core.py", line 1259, in invoke
return _process_result(sub_ctx.command.invoke(sub_ctx))
File "C:\Miniconda3\lib\site-packages\click\core.py", line 1066, in invoke
return ctx.invoke(self.callback, **ctx.params)
File "C:\Miniconda3\lib\site-packages\click\core.py", line 610, in invoke
return callback(*args, **kwargs)
File "C:\Miniconda3\lib\site-packages\click\decorators.py", line 21, in new_func
return f(get_current_context(), *args, **kwargs)
File "C:\rainy\cli.py", line 52, in train
experiment.train(eval_render=eval_render)
File "C:\rainy\experiment.py", line 97, in train
for res in self.ag.train_episodes(self.config.max_steps):
File "C:\rainy\agents\base.py", line 274, in train_episodes
self.store_transition(state, action, *transition[:-1]) # Do not pass info
TypeError: store_transition() takes 4 positional arguments but 6 were given

Thanks for your help!

BTW:
I can run your dqn_cartpole.py like this:
python examples\dqn_cartpole.py train
with no problems.