State Transitions Unlock Temporal Memory in Swarm-Based Reservoir Computing
Date:
Swarms of simple agents provide a physical reservoir substrate: local interactions with motion-induced rewiring. We revisit swarm-based Reservoir Computing (RC) to clarify when swarms exhibit temporal memory and how that capacity scales. Our core result is empirical and reproducible: adding a second internal state with rule-based transitions (dispersed <–> clustered) unlocks substantial memory and yields a linear scaling law with population size. Using a standardized, pure memory-capacity (MC) protocol (i.i.d. input, linear readout, no polynomial terms), we show that two-state swarms achieve high MC that increases with the number of agents N, while single-state swarms remain effectively memoryless under the same protocol. We provide a GPU-ready recipe and validate that the effect is intrinsic to swarm dynamics, not post-processing. We also contextualize the advance by standardizing evaluation and providing a clean, reproducible scaling characterization beyond prior swarm RC reports.