This is interesting from an optimization perspective too. The intermittent big jumps are what you need to get out of local optima. You can use gradient ascent

This is interesting from an optimization perspective too. The intermittent big jumps are what you need to get out of local optima. You can use gradient ascent locally, but once you stop improving, a big change is needed.