While reinforcement learning algorithms converge towards a single policy, it may be useful to generate multiple policies instead of just one.
While reinforcement learning algorithms converge towards a single policy, it may be useful to generate multiple policies instead of just one. This diversity is an indication of what behaviors are within reach. It also helps to cross the reality gap as some policies may transfer better than others and finally the repertoire of generated policies can be used as a set of primitive actions by an upper level policy or planning algorithm.
Language of the library: It is a library in Python, that has been tested for the moment on Linux and MacOS environments.
Additional information: Diversity algorithms like Novelty Search try to cover a “behavior space”. The behavior space is, in general, provided and indicates the space that is worth exploring. It is a projection of the robot trajectory in a smaller dimension space. Quality Diversity algorithms also take into account and thus generate set of solutions that are both diverse and efficient according to a given quality measure. It should be noted that the quality is, in general considered at a local level: the comparison between solutions is done between solutions that are similar, i.e. that are close in the behavior space.
This resource is being used for case studies in the VeriDream project.
- Github repository: https://github.com/robotsthatdream/diversity_algorithms