In situ Distributed Genetic Programming: An Online Learning Framework for Resource Constrained Networked Devices

Abstract

This research presents In situ Distributed Genetic Programming (IDGP) as a framework for distributively evolving logic while attempting to maintain acceptable average performance on highly resource-constrained embedded networked devices.The framework is motivated by the proliferation of devices employing microcontrollers with communications capability and the absence of online learning approaches that can evolve programs for them. Swarm robotics, Internet of Things (IoT) devices including smart phones, and arguably the most constrained of the embedded systems, Wireless Sensor Networks (WSN) motes, all possess the capabilities necessary for the distributed evolution of logic -specifically the abilities of sensing, computing, actuation and communications. Genetic programming (GP) is a mechanism that can evolve logic for these devices using their "native" logic representation (i.e. programs) and so technically GP could evolve any behaviour that can be coded on the device.IDGP is designed, implemented, demonstrated and analysed as a framework for evolving logic via genetic programming on highly resource-constrained networked devices in real-world environments while achieving acceptable average performance.Designed with highly resource-constrained devices in mind, IDGP provides a guide for those wishing to implement genetic programming on such systems. Furthermore, an implementation on mote class devices is demonstrated to evolve logic for a time-varying sense-compute-act problem and another problem requiring the evolution of primitive communications. Distributed evolution of logic is also achieved by employing the Island Model architecture, and a comparison of individual and distributed evolution (with the same and iii iv Abstract slightly different goals) presented. This demonstrates the advantage of leveraging the fact that such devices often reside within networks of devices experiencing similar conditions.Since GP is a population-based metaheuristic which relies on the diversity of the population to achieve learning, many, if not most, programs within the population exhibit poor performance. As such, the average observed performance (pool fitness) of the population using the standard GP learning mechanism is unlikely to be acceptable for online learning scenarios. This is suspected to be the reason why no previous attempts have been made to deploy standard GP as an online learning approach. Nonetheless, the benefits of GP for evolving logic on such devices are compelling and motivated the design of a novel satisficing heuristic called Fitness Importance (FI). FI is population-based heuristic used to bias the evaluation of candidate solutions such that an "acceptable" average fitness (AAF) is achieved while also achieving ongoing, though diminished, learning capacity. This trade off motivated further investigation into whether dynamically adjusting the average performance in response to AAF would be superior to a constant, balanced, performing-learning approach. Dynamic and constant strategies were com...