QDax
QDax copied to clipboard
adaptive-intelligent-robotics
Accelerated Quality-Diversity
Related issues: #181 This PR introduces: A MAP-Annealing repertoire `MAERepertoire` which can be used to run a [CMA MAP-Annealing](https://arxiv.org/pdf/2205.10752) algorithm (CMA-MAE). ## Checks - [ ] a clear description of...
Add MAP-Annealing Repertoire to implement the [Covariance Matrix Adaptation MAP-Annealing](https://arxiv.org/pdf/2205.10752) algorithm (CMA-MAE).
Can't use kwargs x and y, which are positional only for jax>=0.4.28. (or so my computer tells me) e.g. jnp.where(array==array, x=True, y=False) should be jnp.where(array==array, True, False)
The code of [QDAC](https://github.com/adaptive-intelligent-robotics/QDAC) integrates an implementation of [DOMiNO](https://arxiv.org/abs/2205.13521) compatible with QDax. We should add that implementation to the current codebase of QDax.
Related issues: Issue #198 Return metrics after initialisation of repertoires. This PR introduces: - returning of metrics after initialisation of repertoires - updates to notebooks/tests to be compatible with this
As mentioned by @hannah-jan, it’s important to include evaluation metrics of the random initialisation of genotypes so we can see how much the first iterations make a difference.
Add a folder `selectors/`with at least uniform and novelty selectors implemented: - Modify QDPG Emitter, PGA Emitter, MultiEmitter and so on… to use selectors - Add crowding-based selection mechanisms for...
`MAPElitesRepertoire` sampling function should be more modular. For instance, the sampling function can return a pytree of type `MAPElitesRepertoire`, with all the corresponding information (genotypes, fitnesses, descriptors, and so on...)...
Metadata
Owner
Metadata
Accelerated Quality-Diversity