Chris Nota

Pytorch's implementation of `add_hparams` isn't really quite correctly. In particular, it creates a new events file for every write, which is not quite what we want. We should write our...

We are sticking with Tensorboard for Logging. However, we want to make sure that people can easily write and inject their own Loggers. We need to make sure our interface...

Right now, we have a repeated piece of code: ```python hyperparameters = {} for hp in args.hyperparameters: key, value = hp.split('=') hyperparameters[key] = type(agent.default_hyperparameters[key])(value) agent = agent.hyperparameters(**hyperparameters) ``` First of...

E.g.: ``` > all-multiagent-atari pong_v3 ppo ppo Traceback (most recent call last): File "/home/cnota/miniconda3/envs/all/bin/all-multiagent-atari", line 33, in sys.exit(load_entry_point('autonomous-learning-library', 'console_scripts', 'all-multiagent-atari')()) File "/home/cnota/repos/autonomous-learning-library/all/scripts/multiagent_atari.py", line 45, in main presets = { File...

The `State` object itself is unnecessarily limited to `(observation, reward, done, mask)`, etc. It should be possible to construct a `State` with arbitrary keys. Additionally, the state should be nestable,...

I want to change both the names of the run directories, as well as the way the logs get organized. The top-level tabs in tensorboard should include `eval`, `info`, etc.

In short, I want `Preset` to have more controller of the `Experiment` that gets created, in order to support parallel execution, multiprocessing, etc. This gives the user much more control...