This week’s BEACON Researchers at Work blog post is by NC A&T graduate student Patrick Wanko.
Consider today’s car, with its extensive sensors, diagnostics, processing and data storage, and communication capabilities. A far cry from the highly mechanized vehicles of 25 years ago, the functionality and maintenance of such vehicles is heavily inspired by biological organisms. This functionality is being extended beyond the vehicle level to subassemblies and parts – particularly with high value parts for users such as the Department of Defense. These users anticipate using AIT (Automated Identification Technology) with emerging capabilities in data storage, processing, sensor integration, communication, and even energy scavenging. As a result, life cycle management of such parts might be inspired by life cycle process of biological/sociological systems.
Generally, life cycle management for a durable product or part consists of selection of design and operational parameters in a manner to maximize multiple objectives related to performance (functionality, availability and cost). Design parameters are set once at the creation of the product and its subsequent versions. Operating parameters may be changed at any point in the life of the part. Both the minimization of parameter variability and optimization of life cycle performance in given environments are desired. Life cycle performance is driven by design, operating, and environmental parameters with control over the first two. As a result, the domain space for solutions is highly complex and nonlinear, implying metaheuristic search techniques. Optimization of these parameters often involves formation of groups based primarily on operating environment. Life cycle improvement can be made in a decentralized manner by coordination within and across such groups.
As a result, we are looking at ways to modify evolutionary metaheuristic techniques to help improve life cycle management. This research might inform parameter selection approaches for physical systems or assist with virtual life cycle optimization when a performance simulator is available. Two initial concepts that have been developed are modifications to traditional evolutionary metasearch inspired by fish schooling and social networking. In both cases, the intent is not to mimic precisely such behavior. Rather, the intent is to incorporate relevant dynamics into the evolutionary algorithm to provide faster, better solutions for the life cycle management problem – in terms of real-time adjustment of operating parameters and intergenerational adjustment of design parameters.
My research is about what I call Schooling Genetic Algorithm or SGA. It is an enhanced evolutionary algorithm in which the population, divided into tightly formed schools, appears to dynamically search for (individual birth/death gives appearance of school population movement) and eat food within the sea (solution domain), while avoiding, by escaping means, predators.
Photo from http://photography.nationalgeographic.com/photography/photos/schools-fish/
The SGA can be started with individual fish with a bias towards forming schools or pre-formed schools. Once schools are formed, algorithmic parameters (crossover/mutation operators and numbers, selection processes) and to some extent algorithmic processes are influenced by the state of the school. For example, when an attractive location is reached (feeding zone), the emphasis on coordinated movement is lessened and on individual search for good locations is increased. When the food in that area is depleted, the fish need to move to a new location to feed – we can think of the depleted area as having a “penalty function” after a period of feeding that will drive the fish away. When an unattractive area is reached (one with low food availability or high risk of predation), the school shifts to predator avoidance mode which involves faster overall group velocity. Both states might result in bifurcation or convergence of the schools. I would be glad to provide those who are interested with the details about how we implement these concepts.
Initial coding of the system is complete with the dynamic behaviors validated. Now we are looking at solution of sample product life cycle problems and comparison with more traditional approaches. As the work on SGA progresses, so many questions come to mind such as: what is the ideal size of a school with respect to foraging? How can predator avoidance better be performed? Hmm…All fascinating questions – but these are topics for another BEACON post.
For more information about Patrick’s work, please contact him at patrick dot wanko at gmail dot com.
