Mava

Mava copied to clipboard

## What? Epsilon scheduler component and classes are added. ## Why? Epsilon scheduler is needed for exploration in DQN-like algorithms. ## How? * A component is created for epsilon scheduling...

nima-siboni

## What? A counter for the steps of the executor is added. ## Why? The number of steps taken is relevant for some parts of an RL solution, for example...

nima-siboni

Usage documentation is incorrect

4

I'm a first-time Mava user. I'm trying to figure out how to use it. In the [docs homepage](https://id-mava.readthedocs.io/en/latest/#system-implementations) there's a "Usage" section with a code snippet. I'm trying to run...

cool-RR

### Describe the bug When running the robocup example using the image `mava:robocup-latest` I get the following error: ``` [trainer/0] File "/home/app/Mava/mava/systems/tf/maddpg/system.py", line 629, in trainer [trainer/0] _, networks =...

DriesSmit

Support for MultiDiscrete action space

1

Hi there, I have a custom Unity environment for which I wrote a wrapper, and its action space can be described by OpenAI's [MultiDiscrete space](https://github.com/openai/gym/blob/master/gym/spaces/multi_discrete.py#:~:text=%22%22%22This%20represents%20the,5%2C%202%2C%202%20%5D). I've tested my environment and...

dewet99

Resolves Issue #503 ## What? MADQN's executor is implemented. ## Why? The first step towards a complete MADQN system. ## How? The MAPPO design in jax/systems/mappo and its executor are...

nima-siboni

### Feature As the first step to have the first MADQN system, the executor is created. ### Proposal The MAPPO design in jax/systems/mappo and its executor are modified for which...

nima-siboni

### Feature A component for epsilon scheduler, and some common epsilon scheduling logics are implemented ### Proposal * A component is created for epsilon scheduling which has the following callbacks:...

nima-siboni

### Feature Adding step counter to the base system in jax redesign. ### Proposal A new variable is added to the store of the executor. ### Testing Tests are: *...

nima-siboni

### Describe the bug The default executor flattens observations which is undesirable for environments which can benefit from spatial knowledge. ### To Reproduce `observation = executor.store.observation.observation.reshape((1, -1))`

sash-a