Human/Machine Collaborations through Bounded-Error Parameter Estimation Methods
A variety of parametric and non-parametric mathematical models currently describe physical phenomena and systems. These behavioral or predictive models have greatly facilitated the transfer of complex and tedious tasks from humans to machines in several engineering, financial, biological and science applications. However, the mathematical models are neither unique nor perfect in their descriptions of underlying phenomena. In fact, one may argue that any mathematical model can describe a phenomenon given sufficient tolerance for model complexity and error. Bounded-error linear parameter estimation methods, such as set-membership identification , offer an alternative to statistical parameter estimation methods in their estimation of model parameters through a feasible parameter set, in which each element conforms to a strict model structure and error magnitude upper bound. The resulting feasible parameter set which accounts for uncertainties and model approximations is then available for further machine or human processing. This presentation introduces bounding-ellipsoid identification , a deterministic linear parameter estimation method, as a representative example of the broader classes of set-membership and bounded error parameter estimation methods. A discussion of data ranking and sifting, early redundancy detection and outlier rejection properties is presented along with classification, emitter location, tracking and state estimation applications. The potential for participating in adaptive human/machine multi-sensory systems as low-power directionality enhancing components is described. Techniques for joint machine/human participation in identification tasks are suggested. In machine-aided human decisions for instance, machines can apply bounded-error methods to narrow the decision choices before further humans involvement. Human-aided machine decisions can begin by humans defining search areas and error bounds in interactively aiding machines in bounded-error identification or localization tasks. Combining human and machine interactions in identification, decision and search processes through bounded-error methods has the potential of improving each task.