Acknowledgment 4.2.7 Discussion 4.2.4 Generating Plans 4.2.3 Operators 4.2 Planning Troop Allocation 4.2.1 States 4.2.2 Actions 4 Planning and Deployments 4.1 Abstraction of Troop Allocation

We are grateful to many colleagues for discussions that have innuenced this paper. In particular, we would like to thank Philip Dart of DSTO, and members of the Army's Combat Development Directorate for sharing their knowledge and ideas on troop deployments. AI planning techniques were initially developed to solve problems that required the determination of a sequence of actions that could be taken to achieve some goal. A typical problem is one of nding the set of actions a mobile robot must take to move from one location to another. Generally, these type of problems require one to nd a feasible solution, and usually there are no resource constraints on that solution. The problem we are addressing has substantially diierent requirements, an optimal solution and limited resources. Consequently, AI planning techniques need to be adapted and supplemented to address these issues. We have introduced operators that have explicit strategies for dealing with resource limitations. One operator allocated resources to achieve protection then deallocated resources to fullll the resource limitation constraint. Another more sophisticated approach involves nding a protection scheme for an asset that uses resources that are approximately equal to the average amount of available resources per asset. The actual resources that an asset may use are adjusted to reeect the important of the asset and its likelihood of attack. In their basic form AI planning techniques do not address problems that have resource constraints. To use these techniques one must supplement the basic facilities by explicitly encoding in the planning operators strategies that comply with the limited resources constraints. AI planning methods nd feasible solutions to problems. When optimal solutions are required heuristics are needed to direct the search towards nding good solutions. We use an approach that selects an asset's protection scheme based on thèbest' available protection to cost ratio (within the constraints of available resources). Thisìocal' approach to optimization may not nd the optimal solution but produces a solution that is better than one found by chance. After a defensive plan has been found we allocate a score to that plan on the basis of the protection it ooers to the resources it uses. In this way alternative plans may be compared. On substantial problems, backtracking over alternative solutions is a computationally ineeective way of nding an optimum solution. A combination of local optimization and global assessment, as outlined above, appears to be a suitable means for …