From Spikes to Swarms: Evolving Spiking Neural Networks to Create Emergent Swarm Behaviors

Abstract

Drones which can swarm and loiter in a certain area cost hundreds of dollars, but mosquitos can do the same and are essentially worthless. To control swarms of low-cost robots, researchers may end up spending countless hours brainstorming robot configurations and policies to ``organically" create behaviors which do not need expensive sensors and perception.Existing research explores the possible emergent behaviors in swarms of robots with only a binary sensor and a simple but hand-picked controller structure. Even agents in this highly limited sensing, actuation, and computational capability class can exhibit relatively complex global behaviors such as aggregation, milling, and dispersal, but finding the local interaction rules that enable more collective behaviors remains a significant challenge. This is because the collective behavior of a swarm forms from the emergent properties of local agent interactions. Even simple if-then interaction rules can result in complex emergent behaviors, and such rules are deceptively difficult to craft by hand. Efforts to reduce the intuition necessary to craft swarm behaviors largely centers on automating the tuning of controller parameters, or automating search of combinations of human-programmed behaviors. We seek to utilize Spiking Neural Networks (SNNs) as an agent controller, allowing for a wide variety of possible control policies. Provided with a fitness function, the EONS evolutionary algorithm discovers SNNs capable of driving dispersion, aggregation, milling, and diffusion behaviors in a swarm of binary sensing ground vehicles with no modification to the process between these, other than changing the fitness function used. This thesis constitutes three steps towards a generalizable methodology for using SNNs to discover and enact agent-local control policies on binary sensing robots: First, we develop a methodology for training an SNN to replicate the milling behavior in simulation. Then, we apply and refine this methodology for a specific robot model, the TurboPi ground vehicle, and show that the simulated swarm behavior transfers, resulting in milling in a real robot swarm. Finally, we apply this methodology to two more swarm behaviors and metrics from existing literature, aggregation and dispersion, and one behavior with no prior usable metric: diffusion. This represents not only a major step towards lowering the amount of human intuition required to achieve emergent behaviors in swarms, but also presents the ability to reinforce or adjust behaviors more directly than with prior model-based approaches to agent control policies.