Researchers tested whether very small and simply wired reservoir computers can both store and select among multiple chaotic behaviors. They found that these minimal designs can represent more than one chaotic attractor at the same time. However, the same designs usually fail when the system must switch reliably from one attractor to another in response to an external cue.

Reservoir computing is a machine‑learning idea that uses a fixed recurrent dynamical core (the “reservoir”) to transform time series into a high‑dimensional signal, and then trains only a simple readout layer. Echo state networks (ESNs) are a common type of reservoir computer. In most past work the reservoir is large and randomly connected, which is thought to produce rich internal dynamics. The authors instead study deterministic, minimal reservoir topologies — small networks with simple, repeatable wiring patterns.

To test storage and selection of multiple chaotic behaviours, the authors trained single ESNs to learn pairs of chaotic systems. They picked eight benchmark three‑dimensional chaotic systems and formed all 28 unordered pairs from them. They tried two protocols. In the blending technique (BT) a single ESN is trained to generate both systems at once. In the parameter‑aware (PA) approach the ESN is trained so that an external cue should make it switch from one learned system to the other. They evaluated ten different minimal deterministic topologies that have appeared in prior work.

The main result is clear and consistent across this set of experiments. Under the blending protocol, the minimal deterministic reservoirs often succeeded: a single small network could reproduce two coexisting chaotic attractors. But under the cue‑driven protocol, the same minimal designs typically failed to perform robust, cue‑dependent switching. The authors also report that no single topology among the ten consistently outperformed the others for either storage or cue‑dependent selection.