Carl Anderson

Do ants paint trucks better than chickens? Markets versus response thresholds for distributed dynamic scheduling

We studied the dynamic allocation of trucks to paint booths, contrasting two previously proposed schemes in which booths bid against each other for trucks: one based on markets and the other ant-inspired response thresholds. We explore parameter space for several system performance metrics and find that this system is surprisingly easy to optimize and that a number of parameters can be eliminated. We investigate two different threshold reinforcement schemes that give rise to booth specialization and also examine variations of the breaking tie rules that decide among booths when two or more place identical, highest bids for a particular truck. We find that the threshold reinforcement scheme usually used in response threshold applications (local update) fares worse than one with global update of thresholds, and that breaking tie rules previously proposed can be simplified without loss of system performance.

Modeling, quantifying and testing complex aggregate service chains

Service chaining, the act of stringing a sequence of services together to form a new service, is a key element of Web services. However, for Web services to reach its full potential the issue of testing service-chaining at the network level must be resolved. How can one map the microlevel service-service interactions to the macrolevel system performance? For instance, as service chains become longer and more complex, how do they affect end-user quality of service? Focusing on aggregate service chains

Evolutionary testing as both a testing and redesign tool: a study of a shipboard firemain's valve and pump controls

situations in which the user invokes a service that carries out the chain, without the user being aware of the individual services - we tackle these questions using a Java simulation tool to model service chaining, visualize network traffic and quantify service chain complexity. We demonstrate that one can orchestrate very complex service chains in a simple distributed manner and quantify how service chain complexity affects end-user quality of service and network loading.

Methods and systems for interactive evolutionary computing (IEC)

Given the large number of possible fault scenarios, complex control systems are usually impossible to test exhaustively. One promising approach, however, is evolutionary testing (ET) in which a genetic algorithm (GA) evolves critical test situations. We evolved challenges (pipe rupture and water demand) to the valve and pump controls of a shipboard firemain system. ET found minor events that collectively produced significant system failure, and also identified a modification to the ship design that improved system performance. ET can also be used to map the boundary of search space for a given, user-defined criterion. The potential of this powerful, generic technique not just for testing but also for system redesign is thus emphasized.