Computational intelligence for support of military tactical decision making

This thesis seeks to improve the speed and quality of military tactical decision making through the application of computationally intelligent decision support. Fuzzy-Genetic Decision Optimization allows the automated development of military courses of action using three component modules, a combat simulation model, a preference model, and a genetic algorithm. An agent based combat simulation model merges the low-level and generally understood phenomena on the modern battlefield into a higher level and less predictable complex model of land combat in which forces adapt to their ever-changing local situation. An implementation of ordinal preference using fuzzy sets allows a commander to specify his or her preference for the potential outcomes of a military battle. This preference model is easy to elicit using a web-based interface, it accurately reflects the commander's preferences, and it does not require preferential independence among the decision criteria. A genetic algorithm allows the iterative generation and evaluation of potential courses of action. The genetic algorithm generates successive populations of courses of action, the combat simulation model predicts the outcomes of battles using these courses of action, and the fuzzy preference model assigns a fitness value to each outcome. In the final population, fuzzy-genetic decision optimization produces a set of highly fit and potentially different alternative courses of action which accomplish the commander's mission objectives. This system reduces the commander's burden from that of generating possible alternatives and guessing at their possible outcomes to that of selecting from a set of possible alternatives, each accompanied by predicted outcomes. The application of fuzzy-genetic decision optimization in a co-evolutionary scheme during a series of decision experiments allowed military planners to develop tactical courses of action which were robust to changes in the enemy plan and performed significantly better than courses of action developed without automated aids. The hierarchical decomposition of the decision tasks and assignment of those tasks to different agents even further enhanced the performance of plans developed with computationally intelligent aid. Finally, data strip mining is a technique which uses iterative neural network sensitivity analysis to develop predictive models from data sets with a large number of predictors and a relatively small number of observations. Data strip mining extracted from combat simulation data neural network models which predicted the expected number of kills against the enemy and losses from enemy fire given the terrain, enemy situation, and weapons characteristics from different locations on the battlefield. This automated terrain analysis allows unit commanders to quickly identify key battlefield locations which maximize mission goals, given the current terrain and enemy situation. The collection of computational intelligence techniques discussed in this thesis shows great potential for application to decision support on the complex and information rich battlefield of the future.