PSO for Fault-Tolerant Nearest Neighbor Classification Employing Reconfigurable, Analog Hardware Implementation in Low Power Intelligent Sensor Systems

Low-power integrated intelligent sensor systems are of increasing interest for the efficient realization of mobile and distributed realizations. Wireless sensor networks (WSN) are one possible example, where long term sensor vigilance and data acquisition and rather sporadic, brief communication phases occur. Mixed-signal realization, in particular exploiting sub-threshold implementation are interesting, but suffer from susceptibility to environmental and process parameter deviations. Redundancy and reconfiguration based on evolutionary approaches can overcome these problems and raise the yield. This is demonstrated for a behavioral model of a previously implemented one-nearest neighbor (1-NN) reconfigurable mixed-signal classifier. An eye-tracking example will be employed for the case study. Initial classifier prototypes were chosen by an iterative algorithm and the effect of increasing circuit perturbations on performance was systematically investigated. Particle Swarm Optimization (PSO) was applied to readjust or reconfigure the prototypes to restore performance. The experiments show that the adjusting prototypes by PSO can be significantly recovered by the mean of 67% for high perturbation. The yield could be increased considerably, so that PSO and instance based reconfiguration gives broader applicability to low-power mixed-signal circuits.