Mava

Mava copied to clipboard

**Describe the bug** When using my own instance of Reverb and Launchpad I can scale up to 10 executors. However, when I use Mava I can only scale to 6...

DriesSmit

### Feature Investigating using `reverb.TrajectoryDataset` for all recurrent systems. ### Proposal Change `make_dataset_iterator` to use `reverb.TrajectoryDataset` since it appears that TrajectoryDataset works better (faster trainer speed) when using the SequenceAdder....

KaleabTessera

[BUG] Optimizers for critic and policy both backpropagating on observation network

1

**Describe the bug** Both optimizers train the observation network separately which can caused the representation learnt to be unstable. Happens in MAD4PG. **Expected behavior** only one optimizer should work on...

EdanToledo

Memory usage keeps on increasing

4

When trying to run ```bash python examples/meltingpot/train_on_substrates.py ``` The RAM usage continuously keeps on increasing till it's full and gets killed. Is this expected or is there a memory leak...

kinalmehta

**Describe the bug** The method [recurrent_n_step_critic_loss](https://github.com/instadeepai/Mava/blob/develop/mava/components/tf/losses/sequence.py#L26) is broken in its current implementation. This method currently does not update the critic network for the last bootstrap_n timesteps.

DriesSmit

**Please describe the purpose of the feature. Is it related to a problem? ** Adding Noisy Networks as an additional choice for exploration in MADQN - Noisy Networks have seen...

EdanToledo

[chore] instruction for installing mava for developement

2

## What? Added install instruction for people who want to develop for mava to readme. ## Why? I couldn't find this information easily and think it should be in the...

sash-a

### Please describe what `needs` to be maintained? We recently found an error in SMAC where we use `dict.values()` on a dictionary where the order of the returned values are...

DriesSmit

feat: Terminate experiment in case of error

1

## What? Terminate an experiment when one of the nodes has an error. ## Why? This feature will be helpful in case of running multipple experiments, in case one of...

OmaymaMahjoub

### Describe the bug The trainer keeps sampling even if the executors crash. This means that the trainer must somehow sample data that the executor is not generating. This might...

DriesSmit