Automatic exploration of the property space of reservoirs

Reservoir Computing is an efficient implementation of a recurrent neural network for dealing with temporal/sequential data processing. However, in terms of matching reservoir dynamics to tasks, the precise balance of properties (kernel rank, generalization rank, memory capacity, size) of reservoirs will vary. To provide guidance for the generation of reservoirs, we use NSGA-II and MAP-Elites to explore the balance between those properties. We further provide three generation strategies for reservoirs: (a) the optimization of the properties of the random generator, (b) the direct optimization of general purpose reservoir, and (c) a combination of both approaches. We show that each approach can generate reservoirs with different ranges of characteristics, making them thus appropriate for different categories of tasks.