Ant Colony Optimization for Model Checking

Model Checking is a well-known and fully automatic technique for checking software properties, usually given as temporal logic formulas on the program variables. Most of model checkers found in the literature use exact deterministic algorithms to check the properties. These algorithms usually require huge amounts of computational resources if the checked model is large. We propose here the use of Ant Colony Optimization (ACO) to check safety properties of concurrent systems. ACO algorithms are stochastic techniques belonging to the class of metaheuristic algorithms [1] and inspired in the foraging behavior of the real ants. The main idea consists in using artificial ants which simulate this behavior in a graph. Artificial ants are placed on initial nodes of the graph and they jump stochastically from one node to another in order to search the shortest path from the initial node to an objective one. The cooperation among the artificial ants is a key factor in the search. This cooperation is performed indirectly by means of the pheromone trails, which is a model of the chemicals that real ants use for their communication. For more details on ACO algorithms see [2]. We use in this work Max-Min Ant System (MMAS), a kind of ACO algorithm in which the pheromone values are bounded to a given range. For the experiments we use five faulty concurrent systems modelled in PROMELA: basic call2, garp, giop12, marriers4, and phi8. We propose two different versions of the MMAS algorithm. In the first one, called MMAS-b, there is no heuristic information for guiding the ants in the construction phase. We compare this version against Depth-First Search (DFS) and Bread-First Search (BFS) algorithms, which perform also a blind search on the exploration graph and are the most used algorithms in current model checkers. In our secondMMAS algorithm, called MMAS-h, the number of enabled transitions in each state is used as a heuristic information for guiding the ants. We compare this version against the A∗ algorithm using the same heuristic (a novel proposal by Edelkampf, Leue and Lluch-Lafuente [3]). We perform 100 independent runs of the MMAS algorithms to get statistical confidence. In all the cases MMAS algorithms found