![Ant Colony Conclusions
Introduction Background Experiments
Optimization & Future Work
Metaheuristics ACO ACOhg 2007
ACO: Pheromone Update
• Pheromone update
During the construction phase
with
After the construction phase
with
• Trail limits (particular of MMAS)
Pheromones are kept in the interval [τmin, τmax]
London, United Kingdom, July 7-11, 2007 14 / 28](https://image.slidesharecdn.com/gecco-sbse-pdf-120531154342-phpapp02/85/Finding-Safety-Errors-with-ACO-14-320.jpg)
Finding Safety Errors with ACO
- 1. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work 2007 Finding Safety Errors with ACO Enrique Alba and Francisco Chicano London, United Kingdom, July 7-11, 2007 1 / 28
- 2. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Motivation 2007 Motivation • Nowadays software is very complex • An error in a software system can imply the loss of lot of money … … and even human lifes • Techniques for proving the correctness of the software are required • Model checking → fully automatic London, United Kingdom, July 7-11, 2007 2 / 28
- 3. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Explicit State MC Safety Properties State Explosion Heuristic MC 2007 Explicit State Model Checking • Objective: Prove that model M satisfies the property : • SPIN: the property f is an LTL formula Model M LTL formula ¬ f Intersection Büchi automaton s0 !p∨q s0 s5 s3 ∩ s0 s2 = s1 s3 s1 p∧!q q s2 s2 s5 s1 s7 s4 s4 s6 s8 s9 London, United Kingdom, July 7-11, 2007 3 / 28
- 4. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Explicit State MC Safety Properties State Explosion Heuristic MC 2007 Explicit State Model Checking • Objective: Prove that model M satisfies the property : • SPIN: the property f is an LTL formula Model M LTL formula ¬ f Intersection Büchi automaton s0 !p∨q s0 s5 s3 ∩ s0 s2 = s1 s3 s1 p∧!q q s2 s2 s5 s1 s7 s4 s4 s6 s8 s9 London, United Kingdom, July 7-11, 2007 4 / 28
- 5. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Explicit State MC Safety Properties State Explosion Heuristic MC 2007 Explicit State Model Checking • Objective: Prove that model M satisfies the property : • SPIN: the property f is an LTL formula Model M LTL formula ¬ f Intersection Büchi automaton s0 !p∨q s0 s5 s3 ∩ s0 s2 = s1 s3 s1 p∧!q q s2 s2 s5 s1 s7 s4 s4 s6 s8 s9 Using Nested-DFS London, United Kingdom, July 7-11, 2007 5 / 28
- 6. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Explicit State MC Safety Properties State Explosion Heuristic MC 2007 Safety Properties • Safety properties are those expressed by an LTL formula of the form: f=□p where p is a past formula • Finding one counterexample ≡ finding one accepting state s0 Intersection automaton Safety Properties s1 s3 s2 Deadlocks s5 s7 Invariants s4 s6 Assertions … s8 s9 • Classical algorithms for graph exploration can be used: DFS and BFS London, United Kingdom, July 7-11, 2007 6 / 28
- 7. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Explicit State MC Safety Properties State Explosion Heuristic MC 2007 State Explosion Problem • Number of states very large even for small models s0 s1 s3 s2 Memory s5 s7 s4 s6 s8 s9 • Example: Dining philosophers with n philosophers → 3n states 20 philosophers → 1039 GB for storing the states • Solutions: collapse compression, minimized automaton representation, bitstate hashing, partial order reduction, symmetry reduction • Large models cannot be verified but errors can be found London, United Kingdom, July 7-11, 2007 7 / 28
- 8. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Explicit State MC Safety Properties State Explosion Heuristic MC 2007 Heuristic Model Checking • The search for errors can be directed by heuristics using algorithms like A*, IDA*, WA* and Best-First 5 s0 2 2 Heuristic value s1 3 s3 s2 1 s5 0 s7 4 s4 0 s6 6 s8 7 s9 • Different kinds of heuristic functions have been proposed in the past: • Formula-based heuristics • Deadlock-detection heuristics • Structural heuristics • State-dependent heuristics London, United Kingdom, July 7-11, 2007 8 / 28
- 9. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 Metaheuristic Algorithms • Designed to solve optimization problems Maximize or minimize a given function: the fitness function • They can find “good” solutions with a “reasonable” amount of resources Metaheuristic Algorithms Single solution Population London, United Kingdom, July 7-11, 2007 9 / 28
- 10. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 Metaheuristics Classification Single solution Greedy Randomized Iterated Local Variable Adaptive Search Search Neighborhood Procedure Search Tabu Simulated Iterative Guided Local Search Annealing Improvement Search Population Estimation of Evolutionary Scatter Distribution Computation Search Algorithms Ant Colony Particle Swarm Optimization Optimization London, United Kingdom, July 7-11, 2007 10 / 28
- 11. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 Metaheuristics Classification Single solution Greedy Randomized Iterated Local Variable Adaptive Search Search Neighborhood Procedure Search Genetic Algorithms Tabu Simulated Iterative Alba & Troya, 1996 Guided Local Search Annealing Improvement Search Godefroid & Khurshid, 2002, 2004 Population Estimation of Evolutionary Scatter Distribution Computation Search Algorithms Ant Colony Particle Swarm Optimization Optimization London, United Kingdom, July 7-11, 2007 11 / 28
- 12. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 ACO: Introduction • Ant Colony Optimization (ACO) metaheuristic is inspired by the foraging behaviour of real ants • ACO Pseudo-code London, United Kingdom, July 7-11, 2007 12 / 28
- 13. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 ACO: Construction Phase • The ant selects its next node stochastically • The probability of selecting one node Trail k depends on the pheromone trail and the τij heuristic value (optional) of the edge Heuristic i ηij Ni • The ant stops when a complete solution is built m j l k London, United Kingdom, July 7-11, 2007 13 / 28
- 14. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 ACO: Pheromone Update • Pheromone update During the construction phase with After the construction phase with • Trail limits (particular of MMAS) Pheromones are kept in the interval [τmin, τmax] London, United Kingdom, July 7-11, 2007 14 / 28
- 15. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 ACOhg: Motivation • Existing ACO models cannot be applied to the search for errors in concurrent programs The graph is very large, the construction of a complete solution could require too much time and memory In some models the number of nodes of the graph is used for computing the initial pheromone values • We need a new model for tackling these problems: ACOhg (ACO for Huge Graphs) Constructs the ant paths and updates the pheromone values in the same way as the traditional models Allows the construction of partial solutions Allows the exploration of the graph using a bounded amount of memory The pheromone matrix is never completely stored London, United Kingdom, July 7-11, 2007 15 / 28
- 16. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 ACOhg: Huge Graphs Exploration The length of the ant paths is limited by λant What if…? λant Objective node Initial node London, United Kingdom, July 7-11, 2007 16 / 28
- 17. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 ACOhg: Huge Graphs Exploration The length of the ant paths is limited by λant What if…? λant Objective node Initial node Two alternatives Expansion Technique: λant changes After σi steps λant = λant + δl London, United Kingdom, July 7-11, 2007 17 / 28
- 18. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 ACOhg: Huge Graphs Exploration The length of the ant paths is limited by λant What if…? λant Objective node Initial node Two alternatives Missionary Technique: starting nodesλant changes Expansion Technique: for path construction change After σsi steps Second ant + δl λant = λstage Third stage London, United Kingdom, July 7-11, 2007 18 / 28
- 19. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 ACOhg: Pheromones • The number of pheromone trails increases during the search • This leads to memory problems Pheromones • We must remove some pheromone trails from memory Steps Remove pheromone trails τij In the missionary technique, below a given threshold τθ remove all pheromone trails after one stage Pheromones Pheromones Stage Steps Steps London, United Kingdom, July 7-11, 2007 19 / 28
- 20. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Metaheuristics ACO ACOhg 2007 ACOhg: Fitness Function • The fitness function must be able to evaluate partial solutions • Penalties are added for partial solutions and solutions with cycles Partial solution Partial solution Complete solution without cycle with cycle Total penalty Penalty constant for Penalty constant for Path length partial solutions solutions with cycles London, United Kingdom, July 7-11, 2007 20 / 28
- 21. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Models Parameters Results Previous Results 2007 Promela Models • We selected 5 Promela models for the experiments Safety Model LoC States Processes Property giop22 717 unknown 11 Deadlock marriers4 142 unknown 5 Deadlock needham 260 18242 4 LTL formula phi16 34 43046721* 17 Deadlock pots 453 unknown 8 Deadlock * Theoretical result • For all except needham, the states do not fit into the main memory of the computer London, United Kingdom, July 7-11, 2007 21 / 28
- 22. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Models Parameters Results Previous Results 2007 Parameters for ACOhg • The ACOhg model was implemented inside the MALLBA library and then included into the HSF-SPIN model checker Parameter Value Parameter Value Steps 100 ξ 0.5 Colony size 10 a 5 λant 10 ρ 0.8 σs 2 α 1.0 s 10 β 2.0 • Fitness function: length of the path + heuristic + penalty for partial solutions • Two variants: using no heuristic (ACOhg-b) and using it (ACOhg-h) • Machine: Pentium 4 at 2.8 GHz with 512 MB London, United Kingdom, July 7-11, 2007 22 / 28
- 23. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Models Parameters Results Previous Results 2007 Results I: Efficacy • We compare the results of ACOhg algorithms against state-of-the-art model checker algorithms: DFS, BFS, A*, and BF Which algorithm finds errors? • ACOhg algorithms are the only ones that are able to find errors in very large models (marriers20). London, United Kingdom, July 7-11, 2007 23 / 28
- 24. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Models Parameters Results Previous Results 2007 Results II: Details London, United Kingdom, July 7-11, 2007 24 / 28
- 25. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Models Parameters Results Previous Results 2007 Results III: Graphical Comparison • Error trail length vs. memory graph ACOhg algorithms require less memory than BFS They also get shorter (better) error trails than DFS • In general, unlike exhaustive algorithms, ACOhg algorithms keep all the results in a good performance region (high accuracy and efficiency) London, United Kingdom, July 7-11, 2007 25 / 28
- 26. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Models Parameters Results Previous Results 2007 Previous Results with Metaheuristics • GA is the previous metaheuristic algorithm applied to this problem • Godefroid & Khurshid (2002), found errors in phi17 and needham models with GA • To the best of our knowledge, this is the most recent result for this problem using metaheuristics • The results state that ACOhg has higher efficacy and efficiency than GA (even taking into account the differences in the machines) • But we cannot do a fair comparison because the models and the model checkers are different (Verisoft against HSF-SPIN) London, United Kingdom, July 7-11, 2007 26 / 28
- 27. Ant Colony Conclusions Introduction Background Experiments Optimization & Future Work Conclusions and Future Work 2007 Conclusions and Future Work Conclusions • ACOhg is able to outperform state-of-the-art algorithms used nowadays in current model checkers for finding safety errors • ACOhg is able to explore really large concurrent models for which traditional model checking techniques fail • This represents a promising starting point for the use of metaheuristic algorithms in model checking and an interesting subject in SBSE Future Work • Combine ACOhg algorithms with other techniques for reducing the amount of memory: Partial Order Reduction and Symmetry Reduction (in progress) • Include ACOhg into JavaPathFinder for finding errors in Java programs (in progress) • Parallel implementation of ACOhg for this problem (parallel model checkers) London, United Kingdom, July 7-11, 2007 27 / 28
- 28. Finding Safety Errors with ACO 2007 Thanks for your attention !!! Questions? London, United Kingdom, July 7-11, 2007 28 / 28