luciotorre / reply Goto Github PK

Implement the hanoi sample with support for rl-glue

Make sure all code is PEP 8 compatible and passes all lint tests.

Add docstrings for generating documentation using Sphinx.

Implement the rock_paper_scissors sample with support for rl-glue

Add some way to persist a policy across experiments. The idea is being able
to interrupt the experiment without losing all already learned knowledge.

This would also allow for later on loading the learned policy and

- continue learning from an advanced starting point
- just evaluate the policy, to perform a learned behaviour

Before the release, make sure all the code has 100% test coverage.

Implement the gambler sample with support for rl-glue

Change the encoder concept by a more general mapping concept. The storage
takes two mappings as optional parameters: of for the observation space and
one for the action space.

A Mapping translates between a domain space and an image space.

Allow for some way of loading a previously learned policy to use either as
a starting point for continuing learning, or just for evaluation.

Move all encoding/decoding related code into the storage.

Implement the gridworld sample with support for rl-glue

Implement the Sarsa learner along with its corresponding unit tests

Implement a storage that does function aproximation. 
Maybe tiling and kanerva and rbf

Implement eligibility traces for Q (maybe for sarsa too) 

Implement the state_value sample with support for rl-glue

Spaces should be used to define the model. Agent and Environment should use
the same spaces. 

If the space is given, the agent/environment are considered to be
reply-compatible, otherwise, they are rlglue-compatible.

If an agent/environment is reply-compatible, it's attributes can be
accessed by name, otherwise they can only be accessed by position.

Currently rl-glue is a hard dependency. Make this not be like this. rl-glue
should be available if installed, but not required.

When using a rlglue environment, task spec has not the extra parameter
giving the attribute names. Currently in those cases, the spaces ignore all
attributes. 

Make those attributes to be referenceable by index, at least.

Add some way to see/measure how the experiment is going.

Show some measure of performance/error.

Add a complete set of unit tests for the learner.py module

Implement the blackjack sample with support for rl-glue