Recent Advances in Memetic Algorithms
Transcription
Recent Advances in Memetic Algorithms
Recent Advances in Memetic Algorithms Dr. N. Krasnogor Automated Scheduling, Optimisation and Planning Research Group University of Nottingham www.cs.nott.ac.uk/~nxk Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” This talk is based on our new book “Recent Advances in Memetic Algorithms” edited by Hart, Krasnogor & Smith As such my talk today contains only a small snapshot of the book First book that deals exclusively on MAs! Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” So before continuing let me thank the chapter’s authors: • K. Katayama, H. Narihisa, D.A. Pelta, C. Prins, S. Bouchenoua, K. Knodler, J.Poland, P.Merz, N. Krasnogor, J,E. Smith, W.E. Hart, A. Zell, X. Yao, F. Wang, K, Padmanabhan, S. Salcedo-Sanz, S. Areibi, S. Gustafson, A. Sinha, Y. Chen, D. E. Goldberg, E.K. Burke, J.D. Landa-Silva, J. Knowles, D. Corne, D. Wyatt, L. Bull, F. Comellas, R. Gallegos • and of course my co-eds Jim Smith and Bill Hart. • The pictures, graphs and stats in this talk are taken from the book and belong to their authors. • I want also to thank the organizers of this workshop for inviting me to give one of the two plenary talks Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Outline of the Talk PART 1 (25 min) • Memetic Algorithms: the issues involved – – – – – – – – Motivation Lamarckianism vs Baldwinism Diversity Operators Choice Use of Knowledge Specific Considerations for Continuous Domains Initialisation Other Hybridisations Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” PART 2: (15 min) •Showcase Applications: –Maximum Diversity Problem –Protein Structure Prediction –Optimal Engine Calibration –Circuit Partitioning Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” PART 3: (15 min) •Methodologies: –Teams of Heuristics –Cooperative Local Search –On-the-fly Operators Discovery PART 4: (5 min) •Questions Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” PART 1 Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Motivation There are several reasons why it is worthwhile hybridizing: •Complex problems can be decomposable, different subproblems be better solved by different methods: •EA could be used as pre/post processors •Subproblem specific information can be placed into variation operators or into local searchers •In some cases there are exact/approximate methods for subproblems •Well established: generally good black-box optimisers do not exist. This is why successful EAs are usually found in “hybridized form” •EA are good at exploring the search space but find it difficult to zoom-in good solutions •Problems have constraints associated to solutions and heuristics/local search are used to repair solutions found by the EA •If heuristic/local search strategies in MAs are “first class citizens” then a much richer definition of adaptive hybrid metaheuristics is possible: the strategies are generated au pair with the solutions they intend to improve (I.e. self-generating or co-evolving memes) Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” A conservation of competence principle applies: the better one algorithm is solving one specific instance (class) the worst it is solving a different instance (class) [Wolpert et.al.] It cannot be expected that a black-box metaheuristic will suit all problem classes and instances all the time, that is, it is theoretically impossible to have both ready made of-the-shelf general & good solvers. MAs are good algorithmic templates that aid in the balancing act of successfully using a general, of-the-shelf, reusable solvers (EAs) with adds-on instance (class) specific features. Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” The Canonical MA From Eiben’s & Smith “Introduction To Evolutionary Computation” At design time lots of issues arise Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Baldwinism VS Lamarckianism • Lamarkian • traits acquired by an individual during its lifetime can be transmitted to its offspring • e.g. replace individual with fitter neighbour • Baldwinian • traits acquired by individual cannot be transmitted to its offspring • e.g. individual receives fitness (but not genotype) of fitter neighbour Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Baldwin’s “filter” Raw fitness Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Diversity The loss of diversity is specially problematic in MAs as the LS tends to focus excesively in a few good solutions. If the MA uses LS up to local optimae then it becomes important to constantly identify new local optimae If the MA uses partial LS you could still be navigating around the basins of attractions of a few solutions Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Diversity There are various ways to improve diversity (assuming that’s what one wants!): •if the population is seeded only do so partially. •instead of applying LS to every individual choose whom to apply it to. •use variation operators that ensure diversity (assorted) •in the local search strategy include a diversity weigth •modify the selection operator to prevent duplicates •archives •modify the acceptance criteria in the local search: Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Diversity The following modified MC exploits solutions (zooms-in) when the population is diverse. If the population is converged it explores (zooms-out) The temperature T of the MC is defined for each generation as: A new solution is accepted when: when population is diverse T? 0? only accepts improvements when population is converged T ??? accepts both better and worst solutions (explores) Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Operators Choice The choice of LS/Heuristic is one of the most important steps in the design of an MA 1. 2. 3. Local searchers induce search landscapes and there has been various attempts to characterize these. Kallel et.al. and Merz et.al. have shown that the choice of LS can have dramatic impact on the efficiency and effectiveness of the MA Krasnogor formally proved that to reduce the worst case run time of MAs LS move operators must induce search graphs complementary (or disjoint) than those of the crossover and mutation. Krasnogor and Smith have also shown that the optimal choice of LS operator is not only problem and instance dependent but also dependent on the state of the overall search carried by the underlying EA The obvious way to implement 2&3 is to use multiple local searchers within an MA (multimeme algorithms) and we will see that the obvious way of including feedback like that suggested by 1 is to use self-generated multiple local searchers (self-generating MAs aka co-evolving MAs) Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Search landscapes Thanks to P. Merz! Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Multiple Local Searchers Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Use of Knowledge The use of knowledge is essential for the success of a search methods There are essentially two stages when knowledge is used: •At design time: eg, in the form of local searchers/heuristics, specific variation operators, initialization biases, etc. •At run time: •using tabu-like mechanisms to avoid revisiting points (explicit) •using adaptive operators that bias search towards unseen/promising regions of search space (implicit) •creating new operators on-the-fly, eg., self-generating or co-evolving MAs (implicit) With appropriate data-mining techniques we can turn implicit knowledge into explicit and feed it back into the design process! Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Specific Considerations for Continuous Domains There are several factors which makes CD optimisation difficult: •Different scales might be required for local/global searches •It is not always possible to determine when a solution is locally optimal •Long local searchers might be needed to ensure convergence to good optima •Several local searchers exists but they are general methods so they violate the conservation of competence principle. Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Specific Considerations for Continuous Domains The design of CD MAs can be different than the one needed for DD. As there is a need to both do long local searchers and balance it with global search then: •LS is truncated after a number of fitness evaluations •LS is applied sporadically •But these strategies makes it difficult to guarantee convergence To the best of my knowledge the only MAs for CD that have guaranteed convergence to LO are Hart’s Memetic Evolutionary Pattern Search. Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Initialisation Intelligent initialisation of the MA is one of the obvious ways of reusing knowledge: •One does not reinvent the wheel ‘cos existing solutions are reused. •Bias the search mechanism towards more suitable regions of the search space. •Given a CPU budget allocation it might pay to spend some part of the budget in smart initialisations rather than in a pure EA. Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” F Fitness F: fitness after a smart initialization T: time needed by an EA with random initialization to reach F T? T T? Time T? ,T? Time needed by the Intelligent Initialization but remember diversity! If T? < T then it is worth initializing. If T < T? then it is not worth doing it Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Memetic Algorithms: the issues involved Other Hybridisations EA + LS have been used in various other hybridisation schemes: •during the genotype to phenotype mapping prior to evaluation, e.g. in timetabling, scheduling and VRP. •during the mutation or crossover stages, e.g., DPX is a good example of intelligent crossover, and Unger & Moult used a try-best approach for protein folding. Note however that these differ from Xover hill-climbing in that the later does not use problem/instance specific knowledge Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” PART 2 Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications The Maximum Diversity Problem Katayama & Narihisa solve the MDP by means of a sophisticated MA. The MDP: The problem consists in selecting out of a set of N elements, M which maximize certain diversity measure Dij Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications The Maximum Diversity Problem This problem is at the core of various important real-world applications: •Immigration and admission policies •Committee formation •Curriculum design •Portfolio selection •Combinatorial chemical libraries •etc Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications The Maximum Diversity Problem Various features: distinct repair & LS, GRASP for init, diversification phase, accelerated LS. Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications Protein Structure Prediction Estructura primaria Estructura Secundaria Estructura terciaria Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications Protein Structure Prediction Krasnogor, Krasnogor & Smith, Krasnogor & Pelta, Smith have used MAs to study fundamentals of the algorithmics behind PSP in simplified models. Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications Protein Structure Prediction Standard MA template except that Multiple Memes which promote diversity by means of fuzzy rules are used Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications Protein Structure Prediction Membership function for “acceptable” solutions Two distinct “acceptability” concepts Promotes improvements Promotes Diversity Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications Protein Structure Prediction New optimal solutions Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications Optimal Engine Calibration The OEC problem is paradigmatic of many industrial problems. In this problem many combinatorial optimisation problems occur: 1. Optimal Design of Experiments 2. Optimal Test Bed Schedule 3. Look-up Table Calculation Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications Optimal Engine Calibration By P.Merz: Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications Optimal Engine Calibration Standard MA template Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications Circuit Partitioning CP is the task of dividing a circuit into smaller parts. Its an important component of the VLSI Layout problem: is a minimization objective 1. this the division permits the fabrication of circuits physically this is a in constraint distinct components 2. By dividing we conquer: resulting circuits can fit fabrication norms, complexity is reduced 3. Can reduce heat dissipation, energy consumption, etc. Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Showcase Applications Circuit Partitioning From S.Areibi’s chapter: A graphical example Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Intelligent initialisation= GRASP+LS Sophisticated problem specific LS (Fiduccia-Mattheyses) Final LS pass Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” PART 3 Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Methodologies Teams of Heuristics In Burke and Landa Silva’s chapter it is suggested that: The canonical MA template could benefit from including features that Multimeme algorithms other metaheuristics already exploit: included these concepts in recommendations Variable Neighbourhood Search: under this approach a number of different neighbourhood structures are systematically explored, tries to improve the current solution while avoiding poor local optima. A-teams of Heuristics: in A-Teams a set of constructive, improvement and destructive heuristics are asynchronously used to improve solutions. Hyperheuristics: the main concept behind the hyperheuristic is that of managing the application of other heuristics adaptively with the purpose of improving solutions. Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Methodologies Cooperative Local Search Burke and Landa Silva’s chapter observes: In a cooperative local search scheme, each individual carries out its own LS. When an individual gets stuck it ask for the cooperation of the population in order to find something to do to get unstuck and continue the search from another position in the solution space. The results achieved by each individual may be different at different times and this encourages diversity within the population Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Methodologies Cooperative Local Search Cycle of each individual in pop The search cycle of each individual begins Cooperation mechanism Gets stuck sharing moves, parts, centralized control, etc Finds something to do. Gets unstuck Note that this differs from teams of heuristics in that here the cooperation is made explicit Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Methodologies On-the-fly operators discovery All the previous methodologies clearly benefits the end user as they have been shown to provide improvements in robustness, quality, etc. But what do we do if we do not have, or don’t know, good heuristics which could be used by,eg., A-teams, VNS or CLS? Also, why don’t we use the information the algorithm produces to better understand and make explicit new knowledge of the search landscape capturing this knowledge in new operators? Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Methodologies On-the-fly operators discovery Two alternatives: 1. Off-line: Whitley and Watson did it successfully for TS, and Kallel et al for other methods. 2. In-line: Krasnogor, Krasnogor & Gustafson, Smith for MAs The problem with 1 is that it takes ages to data mine the data and make it reusable, although it is STILL worth doing it. And what about 2? Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Methodologies On-the-fly operators discovery Canonical MA cycle Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Methodologies On-the-fly operators discovery Self-Generating/Co-evolving MAs Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Methodologies On-the-fly operators discovery •Inheritance: an agent inherits the meme of the most successful of its parents There are various processes that guide the Agent’s cultural evolution of local search strategies: •Imitation: an agent imitates a successful non-genetically-related individual •Innovation: an agent blindly (i.e.randomly) change its meme •Mental Simulation: an agent purposely (e.g. hill-climbs to ) improve its meme Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” From Krasnogor & Gustafson chapter Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” From Smith chapter Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Conclusions (I) •There is much more in MA that meets the eye. Its not a simple matter of ad-hoc putting LS somewhere in the EA cycle. •Just a small space of the architectural space of MAs has been explored and we don’t know yet why a given architecture performs well/bad in a specific problem (see my thesis) •People usually use one “silver bullet” LS. That’s fine if that SB exists. However when it does not exist use multimeme algorithms, or other heuristics teams/cooperative algorithms as lots of simple heuristics can synergistically do the trick. Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Conclusions (II) •ADAPT: the search process is dynamic and your method should detect and adapt to changing circumstances. Adaptation is not too expensive or complex to code! •Carefully consider how your variation operators interact with LS •Ideam for Baldwinian VS Lamarckianism •Understand that the fitness landscape explored by your MA is not a one-operator landscape but the results of the superposicion with interference of varios landscapes. Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Conclusions (III) •Use more expresive acceptance criteria in your local search, eg., fuzzy criteria •If you don’t know what operators to apply let the the MA find it for you by some Self-Generating mechanism, e.g., co-evolution. •Self-Generating mechanisms are a great niche for GPers! FINALLY: check out the literature, almost surely you will find MAs. among the best success stories in applications to real world probs! Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics” Thank you! Questions? Recent Advances in Memetic Algorithms Plenary Talk at the “EU/ME Workshop: Design and Evaluation of Advanced Hybrid Metaheuristics”