On Predictions for Dynamic, Self-Adaptive Techniques in DynDCSP´s

In many situations, the distributed constraint satisfaction problem provides a useful way to model real-world problems and apply agent-based solutions. However, more often than not, the problems that we wish to model change over time, and as such map better to dynamic DCSP´s than to static instances. Until recently, algorithms to solve dynamic DCSP´s (DynDCSP) have been developed and evaluated purely on the basis of empirical testing, and environmental rates of change have only been addressed in an ad-hoc manner. In this study, we leverage a recent theoretical thermodynamic model for DynDCSP´s to generate performance predictions that guide parameter choice in the Distributed Stochastic Algorithm (DSA) automatically. The technique is applicable to any algorithm, and allows it to adjust its parameters according to observed environmental characteristics with the goal of maximizing its solution quality over time. We examine the effectiveness of using predicted parameter values from static experimentation to choose appropriate parameters in dynamic scenarios, and consider the impact of different modes of environmental estimation on effective parameter choice.