Meta-optimalizace solveru systému generálního a hlavního klíče

Abstract

Many practical problems are NP-Hard (or NP-Complete), for which best algorithms may guarantee only exponential worth-case time complexity, if P != NP. Typically, these algorithms can not do better, but several types of heuristics can be invented into algorithms to improve their performance, without any guarantee for better worth-case complexity. Lock-chart solving is a NP-Hard problem. It is specified by a lock system - a set of keys and locks installed in one or several buildings, for each key customer decides which doors it should be able to open. Lock-Chart problem is formulated by these requirements. For lock-chart solving many different combinatorial optimization algorithms and techniques can be used. One of which is backtracking algorithm. This algorithm can use different pruning technique and heuristics. Heuristics are typically chosen randomly on each backtrack, with some participation of random shakes technique. This master thesis presents approach of meta-heuristic optimization for Lock-chart problem Solver based on Machine Learning techniques. Main efforts of this thesis are: patterns and dependencies between heuristics and lock-chart problems were discovered, Decision Tree based heuristic selection system was created and included into the current backtracking algorithm, performance improvement was measured by two experiments. Proposed Decision tree based heuristic selection approach outperformed the random heuristic selection approach in 92 problem instances out of 113.

Many practical problems are NP-Hard (or NP-Complete), for which best algorithms may guarantee only exponential worth-case time complexity, if P != NP. Typically, these algorithms can not do better, but several types of heuristics can be invented into algorithms to improve their performance, without any guarantee for better worth-case complexity. Lock-chart solving is a NP-Hard problem. It is specified by a lock system - a set of keys and locks installed in one or several buildings, for each key customer decides which doors it should be able to open. Lock-Chart problem is formulated by these requirements. For lock-chart solving many different combinatorial optimization algorithms and techniques can be used. One of which is backtracking algorithm. This algorithm can use different pruning technique and heuristics. Heuristics are typically chosen randomly on each backtrack, with some participation of random shakes technique. This master thesis presents approach of meta-heuristic optimization for Lock-chart problem Solver based on Machine Learning techniques. Main efforts of this thesis are: patterns and dependencies between heuristics and lock-chart problems were discovered, Decision Tree based heuristic selection system was created and included into the current backtracking algorithm, performance improvement was measured by two experiments. Proposed Decision tree based heuristic selection approach outperformed the random heuristic selection approach in 92 problem instances out of 113.