Metazoans gather information from their environments and respond in predictable ways. These computational tasks are achieved with neural networks of varying complexity. Their performance must be reliable over an individual's lifetime while dealing with the shorter lifespan of cells and connection failure-thus rendering ageing a relevant feature. How do computations degrade over an organism's lifespan? How reliable can they remain throughout? We tackle these questions with a multi-objective optimization approach. We demand that digital organisms equipped with neural networks solve a computational task reliably over an extended lifespan. Neural connections are costly (as an associated metabolism in living beings). They also degrade over time, but can be regenerated at some expense. We investigate the simultaneous minimization of both these costs and the computational error. Pareto optimal trade-offs emerge with designs displaying a broad range of solutions: from small networks with high regeneration rate, to large, redundant circuits that regenerate slowly. The organism's lifespan and the external damage act as evolutionary pressures. They improve the exploration of the space of solutions and impose tighter optimality constraints. Large damage rates can also constrain the space of possibilities, forcing the commitment of organisms to unique strategies for neural systems maintenance.
Keywords: ageing; artificial neural networks; multi-objective optimization; regeneration.