search

ci-group / revolve2

A python library for optimization, geared towards modular robots and evolutionary computing.
https://ci-group.github.io/revolve2
GNU Lesser General Public License v3.0
18 stars 49 forks source link

Novelty search #99

Open karinemiras opened 2 years ago

surgura commented 2 years ago

Is this a specific algorithm? Or does this simply mean selection based on phenotypical features? In the latter case we would simply need to provide the ability to measure modular robots(#97 )? The existing EA optimizer might alredy provide a framework for the evolution process.

karinemiras commented 2 years ago

Neither one nor another, but a hybrid. This is an optimization towards diversity: the dimensions should be a parameter - the user can choose which traits (behavioral, phenotypical, genotypical, etc) to use for defining diversity. It requires a measure of diversity, but also an archive to make sure the search won't go in circles. You can read the original paper: https://ieeexplore.ieee.org/document/6793380 Or take a look at the current code: https://github.com/ci-group/revolve/blob/plasticoding/pyrevolve/evolution/population.py relevant methods of the file script: def calculate_novelty, def calculate_distances, def collect_measures, def update_archive

surgura commented 2 years ago

Let's create a new example and meanwhile implement features in revolve we need for that