Improving our knowledge of metaheuristic approaches for cell suppression problem

1. Improving our knowledge of metaheuristic approaches for cell suppression problem Andrea Toniolo Staggemeier Alistair R. Clark James Smith Jonathan Thompson

2. Content • Introduction • Business Problem • Comparison of Existing Methods in Tau-Argus • Proposed Solutions • Experiments Design • Parameter Optimisation • Operator Approach • Table Safety • Hierarchical Tables • Issues • External bounds in attacker model and tight intervals set in Tau-Argus • Upper, Lower, and Sliding Protection levels set in Tau-Argus • The Incremental Attacker Heuristic • Conclusion

3. Introduction • Business Problem • Various methodologies for Cell Suppression • Large instances • Hierarchical classifications • Multidimensional data • Frequency data • Computational time • Comparison of Existing Methods in Tau-Argus • Means of assessing fair existing methods and new methods • Proposed Solutions • Metaheuristic vs.. Mathematical Programming – Pros and Cons

4. Experiments Design • Parameter Optimisation • GRASP • ANT Colony • EAs • Operator Approach • Better operators GREEDY and Descent Method • Alternative Mutation operators EA • Table Safety • Relaxed feasibility check vs. strict feasibility check • Improved heuristic for checking table safety • Hierarchical Tables • Effect hierarchical data structure has on problem/algorithms

5. Issues • External bounds in attacker model and tight intervals set in Tau-Argus • Attackers knowledge assumptions • How upper and lower bounds are set • Upper, Lower, and Sliding Protection levels set in Tau-Argus • Strictly less than constraints • Special issue on frequency data • What happened with Sliding protection information? • The Incremental Attacker Heuristic • Who should we give protection: primary, of course, but if the secondary cells added are not sufficient protected can I disclose a primary by consequence?

6. Conclusion • ACO best method but too slow. Could be implemented in parallel • Greedy heuristics promising, good solution quality, extremely quick and easy to adapt • Large dataset, (4000 * 12), solved in 328 seconds • Over 80% of solutions strictly feasible • Can adapt to hierarchical tables • For datasets that were not strictly feasible, feasibility test resulted in cost increasing by an average of 57%

7. Which is the best meta-heuristic for solving the relaxed cell suppression problem? GRASP What are best parameters? How robust are they? Effective on all datasets How time consuming is the feasibility test? Varies according to problem size How effective is a “relaxed” feasibility test? Very effective in some cases Does priming the feasibility test IP with a good starting solution produce better results than the feasibility test IP alone? In most cases What is the quality of results on different sorts of datasets? Varies, no pattern but seemingly worse on larger data. More time consuming on datasets with more sensitive cells / less zero cells. How effective is the general method when applied to the heuristic problem? Very - solutions equal lower bound in most cases.

8. Genetic algorithms: • Can reliably improve on heuristic solutions • Outperform Local Search approaches • Have identified suitable operators/parameters • BUT • Approach based on treating sensitive cells individually is too slow for large tables • Need to do more research into grouping approach • Seems more scalable • Good reasons to believe it may give better solutions

9. PhD research underway to look at the problems found and try to develop alternative methods using metaheuristic approaches based on this work.