Memetic Algorithms Dr. N. Krasnogor Interdisciplinary Optimisation Laboratory Automated Scheduling, Optimisation and Planning Research Group School of Computer Science & Information Technology University of Nottingham www.cs.nott.ac.uk/~nxk Ben Gurion University of the Negev – 2nd May 2010 Outline of the Talk – Evolutionary Algorithms Revisited – MAs’ Motivation (General Versus Problem Specific Solvers) – MAs design issues – A Few Exemplar Applications – Related Methods and Advanced Topics – Putting it all Together – Conclusions, Q&A Ben Gurion University of the Negev – 2nd May 2010 Based on: o N. Krasnogor. Handbook of Natural Computation, chapter Memetic Algorithms. Natural Computing. Springer Berlin / Heidelberg, 2009. o J. Bacardit and N. Krasnogor. Performance and efficiency of memetic pittsburgh learning classifier systems. Evolutionary Computation, 17(3), 2009. o Q.H. Quang, Y.S. Ong, M.H. Lim, and N. Krasnogor. Adaptive cellular memetic algorithm. Evolutionary Computation, 17(3), 2009. o N. Krasnogor and J.E. Smith. Memetic algorithms: The polynomial local search complexity theory perspective. Journal of Mathematical Modelling and Algorithms, 7:3-24, 2008. o M. Tabacman, J. Bacardit, I. Loiseau, and N. Krasnogor. Learning classifier systems in optimisation problems: a case study on fractal travelling salesman problems. In Proceedings of the International Workshop on Learning Classifier Systems, volume (to appear) of Lecture Notes in Computer Science. Springer, 2008. o N. Krasnogor and J.E. Smith. A tutorial for competent memetic algorithms: model, taxonomy and design issues. IEEE Transactions on Evolutionary Computation, 9(5):474- 488, 2005. o W.E. Hart, N. Krasnogor, and J.E. Smith, editors. Recent advances in memetic algorithms, volume 166 of Studies in Fuzzyness and Soft Computing. Springer Berlin Heidelberg New York, 2004. ISBN 3-540-22904-3. o N. Krasnogor. Self-generating metaheuristics in bioinformatics: the protein structure comparison case. Genetic Programming and Evolvable Machines, 5(2):181-201, 2004. o N.Krasnogor and S. Gustafson. A study on the use of “self-generation” in memetic algorithms. Natural Computing, 3(1):53 - 76, 2004. o M. Lozano, F. Herrera, N. Krasnogor, and D. Molina. Real-coded memetic algorithms with crossover hill-climbing. Evolutionary Computation, 12(3):273-302, 2004. All material available at www.cs.nott.ac.uk/~nxk/publications.html Ben Gurion University of the Negev – 2nd May 2010 Evolutionary Computation Most Important Metaphors Evolution Problem Solving Environment Problem Individual Candidate/Feasible Solution Fitness Solution quality, i.e. objective value Natural Selection Simulated Pruning of bad solutions Fitness survival and reproduction likelihood Objective value chances of generating related (but not necessarily identical) solutions A.E. Eiben and J.E. Smith, Introduction to Evolutionary Computing Hybridisation with other techniques: Memetic Algorithms Ben Gurion University of the Negev – 2nd May 2010 Simulated Evolution Consists of: • The population contains a diverse set of individuals (i.e. solutions to an optimisation problem) • Those features that make solutions good under a specific objective function tend to proliferate • Variation is introduced into the population by means of mutation, crossover and, for MAs, local search. • Selection culls the population throwing out bad solutions and keeping the most promising ones Ben Gurion University of the Negev – 2nd May 2010 The Metaphor of “Adaptive landscape” (Wright, 1932) (1) • Solutions (i.e. individuals) are represented by n properties + its quality value. • One can imagine the space of all solutions and their quality values represented as a graph (i.e. landscape) in n+1 dimensions. • In this metaphor, a specific individual is seen as a point in the landscape. • Each point in the landscape will have neighbouring points, which are those solutions that are somehow related to that point implying a neighbourhood. • It is called “adaptive” landscape because the quality value function, F(), depends on the features and properties of the individual and hence the better those are the higher(smaller) the value. Ben Gurion University of the Negev – 2nd May 2010 An example: Maximisation Problem with HillClimber (2) Objective Value a solutions Solutions have 2 features Prop 2 Prop 1 Ben Gurion University of the Negev – 2nd May 2010 An example: Maximisation Problem with EA (3) Prop 2 Prop 1 Ben Gurion University of the Negev – 2nd May 2010 Evolutionary Algorithms in Context • There are several opinions about the use of metaheuristics in optimisation • For the majority of problems a specific algorithm could: – work better than a generic algorithm in a large set of instances , – but it can be very limited on a different domain. – don’t work too good for some instances. • One important research challenge is : – to build frameworks that can be robust across a set of problems delivering good enough/cheap enough/soon enough solutions. – to a variety of problems and instances. Ben Gurion University of the Negev – 2nd May 2010 Outline of the Talk – Evolutionary Algorithms Revisited – MAs’ Motivation (General Versus Problem Specific Solvers) – MAs design issues – A Few Exemplar Applications – Related Methods and Advanced Topics – Putting it all Together – Conclusions, Q&A Ben Gurion University of the Negev – 2nd May 2010 method performance on problems EAs as problem solvers: The pre-90s view The For T for problem solving the Rollroyce for problem P specific method metaheuristic method random search P scale of all problems A.E. Eiben and J.E. Smith, Introduction to Evolutionary Computing Hybridisation with other techniques: Memetic Algorithms Ben Gurion University of the Negev – 2nd May 2010 Evolutionary Algorithms and domain knowledge • Fashionable after the 90’s: to add problem specific information into the EAs by means of specialized crossover, mutation, representations and local search • Result: The performance curve deforms and – makes EAs better in some problems, – worst on other problems – amount of problem specific is varied. A.E. Eiben and J.E. Smith, Introduction to Evolutionary Computing Hybridisation with other techniques: Memetic Algorithms Ben Gurion University of the Negev – 2nd May 2010 Michalewicz’s Interpretation method performance on problems EA 4 EA 2 EA 3 EA 1 P scale of all problems A.E. Eiben and J.E. Smith, Introduction to Evolutionary Computing Hybridisation with other techniques: Memetic Algorithms Ben Gurion University of the Negev – 2nd May 2010 So What Are Memetic Algorithms? MAs are carefully orchestrated interplay between (stochastic) global search and (stochastic) local search algorithms N. Krasnogor. Handbook of Natural Computation, chapter Memetic Algorithms. Natural Computing. Springer Berlin / Heidelberg, 2009 Ben Gurion University of the Negev – 2nd May 2010 So What Are Memetic Algorithms? Adding Domain Knowledge to EAs Memetic Algorithms (MAs) were originally inspired by: • Models of adaptation in natural systems that combine evolutionary adaptation of population of individuals (GAs) WITH • Individual learning (LS) within a lifetime (others consider the LS as development stage). Learning took the form of (problem specific) local search PLUS • R. Dawkin’s concept of a meme which represents a unit of cultural evolution that can exhibit refinement, hence the local search can be adaptive. BUT WHY? Ben Gurion University of the Negev – 2nd May 2010 An example: Maximisation Problem with EA (3) Prop 2 Prop 1 Ben Gurion University of the Negev – 2nd May 2010 The Canonical MA At design From Eiben’s & Smith “Introduction To Evolutionary Computation” Ben Gurion University of the Negev – 2nd May 2010 time lots of issues arise Memetic Algorithms: the issues involved Motivation There are several reasons why it is worthwhile hybridizing: •Complex problems can be partially 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 theoretical results: generally good black-box optimisers do not exist. This is why successful EAs are usually found in “hybridized form” with domain knowledge added in •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 one can raise the level of generality of the algorithm without sacrificing performance by letting the MA choose which local search to use. Ben Gurion University of the Negev – 2nd May 2010 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 for all problems. MAs and Hyperheuristics are good algorithmic templates that aid in the balancing act of successfully & cheaply using general, of-theshelf, reusable solvers (EAs) with adds-on instance (class) specific features. Ben Gurion University of the Negev – 2nd May 2010 What happens inside an MA? This is my solution to your problem I used strategy X Memetics Population of agents When I grow up I’ll need to decide whose problem solving strategy to use Evolutionary Algorithm Offspring Ben Gurion University of the Negev – 2nd May 2010 Outline of the Talk – Evolutionary Algorithms Revisited – MAs’ Motivation (General Versus Problem Specific Solvers) – MAs design issues – A Few Exemplar Applications – Related Methods and Advanced Topics – Putting it all Together – Conclusions, Q&A Ben Gurion University of the Negev – 2nd May 2010 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 Ben Gurion University of the Negev – 2nd May 2010 Ben Gurion University of the Negev – 2nd May 2010 Baldwin’s “filter” Raw fitness Ben Gurion University of the Negev – 2nd May 2010 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 Ben Gurion University of the Negev – 2nd May 2010 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: Ben Gurion University of the Negev – 2nd May 2010 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 goes to infinity accepts both better and worst solutions (explores) Ben Gurion University of the Negev – 2nd May 2010 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) Ben Gurion University of the Negev – 2nd May 2010 Memetic Algorithms: the issues involved Fitness Landscapes Thanks to P. Merz! Ben Gurion University of the Negev – 2nd May 2010 Multiple Local Searchers Ben Gurion University of the Negev – 2nd May 2010 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 (implicit) • 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 (explicit) With appropriate data-mining techniques we can turn implicit knowledge into explicit and feed it back into the design process! (Deb Calls this Innovisation) Ben Gurion University of the Negev – 2nd May 2010 Outline of the Talk – Evolutionary Algorithms Revisited – MAs’ Motivation (General Versus Problem Specific Solvers) – MAs design issues – A Few Exemplar Applications – Related Methods and Advanced Topics – Putting it all Together – Conclusions, Q&A Ben Gurion University of the Negev – 2nd May 2010 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 Ben Gurion University of the Negev – 2nd May 2010 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 Ben Gurion University of the Negev – 2nd May 2010 Showcase Applications Protein Structure Prediction by us Primary Structure Tertiary Structure Ben Gurion University of the Negev – 2nd May 2010 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. Ben Gurion University of the Negev – 2nd May 2010 Showcase Applications Protein Structure Prediction Standard MA template except that Multiple Memes which promote diversity by means of fuzzy rules are used Ben Gurion University of the Negev – 2nd May 2010 Showcase Applications Protein Structure Prediction Membership function for “acceptable” solutions Two distinct “acceptability” concepts Promotes Diversity Promotes improvements Ben Gurion University of the Negev – 2nd May 2010 Showcase Applications Protein Structure Prediction New optimal solutions Ben Gurion University of the Negev – 2nd May 2010 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 Ben Gurion University of the Negev – 2nd May 2010 Showcase Applications Optimal Engine Calibration By P.Merz: Ben Gurion University of the Negev – 2nd May 2010 Showcase Applications Optimal Engine Calibration Standard MA template Ben Gurion University of the Negev – 2nd May 2010 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 ainconstraint distinct components 2. By dividing we conquer: resulting circuits can fit fabrication norms, complexity is reduced 3. Can reduce heat dissipation, energy consumption, etc. Ben Gurion University of the Negev – 2nd May 2010 Showcase Applications Circuit Partitioning From S.Areibi’s chapter: A graphical example Ben Gurion University of the Negev – 2nd May 2010 Showcase Applications The Maximum Diversity Problem Various features: distinct repair & LS, GRASP for init, diversification phase, accelerated LS. Ben Gurion University of the Negev – 2nd May 2010 Outline of the Talk – Evolutionary Algorithms Revisited – MAs’ Motivation (General Versus Problem Specific Solvers) – MAs design issues – A Few Exemplar Applications – Related Methods and Advanced Topics – Putting it all Together – Conclusions, Q&A Ben Gurion University of the Negev – 2nd May 2010 Related Methodologies Teams of Heuristics 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. Ben Gurion University of the Negev – 2nd May 2010 Methodologies Cooperative Local Search (Landa Silva & Burke) 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 Ben Gurion University of the Negev – 2nd May 2010 Methodologies Off-line & In-line 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? Ben Gurion University of the Negev – 2nd May 2010 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. Tabacman et al demonstrated it for TSP 2. In-line: Krasnogor, Krasnogor & Gustafson, J.E. Smith and others for MAs on PSP, PSC & other problems Ben Gurion University of the Negev – 2nd May 2010 • Strategic Goals: Objectives – To move away from a close-system optimisation scenario to an open-andembodied- optimisation scenario – To integrate Datamining and Classification techniques within the life cycle of optimisation problem solvers • Tactical Goals: – To use Learning Classifier Systems / Genetic Programming / Etc to learn features from instances of an optimisation problem – Provide a TSP Solver as a proof of concept: • The Fractal TSP family of problems is used • GAssist is trained to distinguish between different members of the family • The predictions are fed into a naive TSP solving heuristic Ben Gurion University of the Negev – 2nd May 2010 Off-line pattern discovery with Learning Classifier Systems • Machine Learning technique • Classifies a set of (previously unseen) input data based on (previously learnt) classification rules • “IF condition THEN action” rules (other semantics are possible) • Use (some flavour of) Evolutionary Algorithms Ben Gurion University of the Negev – 2nd May 2010 Traveling Salesman Problem(TSP) • Goal: Given a set of cities’ coordinates, find the Hamiltonian cycle of minimum cost (search on a Kn graph, n=# of cities) • One of the most studied combinatorial optimization problems • NP-hard problem Ben Gurion University of the Negev – 2nd May 2010 Fractal TSP • A family of TSP 2D euclidean instances described by L-Systems rules • Coordinates are generated while the L-System rules are being applied • Sequence of coordinates describes the optimal solution for the underlying TSP • Instances are built by iterating over the rewriting rules (axioms) of an L-System • Tour of order n is obtained by iterating the L-system n times • Higher orders give rise to larger and more intrincate city distributions: #cities=f(n) Ben Gurion University of the Negev – 2nd May 2010 Fractal TSP David Tour Koch Mpeano MNpeano Ben Gurion University of the Negev – 2nd May 2010 No Free Lunch Theorem • In general: No algorithm is unsurpassed in efficiency for all possible inputs • For Fractal TSP: No heuristic gives the optimal solution for all families of FTSPs • Fractal TSP Solvers – Mpeano Nearest Neighbour & Multiple Fragment – MNPeano Multiple Fragment – Koch Furthest Addition From Minimal Convex Hull Theoretically proved to be optimally solved by Ben Gurion University of the Negev – 2nd May 2010 Fractal TSP • By Construction the optimal tour is known • Hence Fractal L-systems are an ideal test bed for TSP methods as: – Scale to as large instances as desired – Instances with features at all scales, e.g. clusters of cities, bridges, etc. – For each instance the optimal tour is known – It is possible to mathematically demonstrate (in some cases) whether a family of instances (for all n) is solvable by a given heuristic Ben Gurion University of the Negev – 2nd May 2010 An LCS Approach to FTSP Optimisation • Described by two interacting algorithms – Instance Classifier (IC) (point 2.1 previous slide) – Greedy Ramdomised Algorithm (GRA) (point 2.2 previous slide) Ben Gurion University of the Negev – 2nd May 2010 Greedy Randomised Algorithm Assuming we have a good classifier for deciding edges inclusion in optimal tour: • Iterates over all of the edges of the complete graph of the TSP instance • Classify an edge. If it belongs to optimal tour include it, reject otherwise • Does not ensure optimality • Obtains quick solutions Ben Gurion University of the Negev – 2nd May 2010 FTSP Instance Classifier • Requires – C: a set of coordinates samples derived from FTPS instances – F = Global Statistics features to be computed for C • Returns: – R: a rule set that classifies a set of Global Statistics features into one of the Fractal TSP tours – Gives the prediction made by R from the data F Ben Gurion University of the Negev – 2nd May 2010 Global Statistics Model • Features: – Number of cities in the sample – Maximum X and Y components – Average spread from X and Y axis center • Expresses a general signature of the shape of the curve in a 2dimensional space. • The Spread of a coordinate from a given axis center computes "how far" the corresponding component is. – Average of nearest neighbors – Features are computed after normalizing the coordinates, with the curve meant to fit a 1x1 unit square Ben Gurion University of the Negev – 2nd May 2010 Triplets Model • Accuracy increases for higher orders Ben Gurion University of the Negev – 2nd May 2010 David Tour Ben Gurion University of the Negev – 2nd May 2010 Koch Tour Ben Gurion University of the Negev – 2nd May 2010 MPeano Ben Gurion University of the Negev – 2nd May 2010 Observations • Accuracy over a random coin classication increases as the number of coordinates grow • As the problem becomes more complex, the proposed system improves its performance over a random solution Ben Gurion University of the Negev – 2nd May 2010 Observations • Simple algorithm structure • Rule sets obtained with basic features, in fact the most obvious. There are plenty of geometrical features that could be tried out. • A more complex algorithm coupled with refined predictions could guide the creation of high performing solutions for the TSP. Ben Gurion University of the Negev – 2nd May 2010 Observations • We can correctly classify, using LCS, a subset of cities into one of four fractal TSP families – Helps deciding which heuristic to use for solving them • LCS might be able to learn how to distinguish betwen parts of an optimal tour • The GAssist LCS provides insight on why it chooses one edge over another to be included in the optimal tour as its rules are human-readable Ben Gurion University of the Negev – 2nd May 2010 Methodologies In-line operators discovery Canonical MA cycle Adapted from Durham, W.: Coevolution: Genes, Culture and Human Diversity. Stanford University Press (1991) Ben Gurion University of the Negev – 2nd May 2010 Methodologies On-the-fly operators discovery Self-Generating/Co-evolving Mas Adapted from Durham, W.: Coevolution: Genes, Culture and Human Diversity. Stanford University Press (1991) Ben Gurion University of the Negev – 2nd May 2010 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 Ben Gurion University of the Negev – 2nd May 2010 An Example with The MAX-CMO Problem Goal is to maximise number of cycles of size = 4 without crossing edges Ben Gurion University of the Negev – 2nd May 2010 Random Instance Generator Number of vertices Density of random edges Number of regular edges pattern 1-7 1—7 N Patterns Pattern probability 3-7-11 3-7-11 Ben Gurion University of the Negev – 2nd May 2010 Ben Gurion University of the Negev – 2nd May 2010 A Real World Instance Ben Gurion University of the Negev – 2nd May 2010 Chunks & Templates Patterns & Grammars Grammars Patterns Ben Gurion University of the Negev – 2nd May 2010 From Krasnogor & Gustafson chapter Ben Gurion University of the Negev – 2nd May 2010 Ben Gurion University of the Negev – 2nd May 2010 Ben Gurion University of the Negev – 2nd May 2010 Ben Gurion University of the Negev – 2nd May 2010 Ben Gurion University of the Negev – 2nd May 2010 From Smith chapter Ben Gurion University of the Negev – 2nd May 2010 Outline of the Talk – Evolutionary Algorithms Revisited – MAs’ Motivation (General Versus Problem Specific Solvers) – MAs design issues – A Few Exemplar Applications – Related Methods and Advanced Topics – Putting it all Together – Conclusions, Q&A Ben Gurion University of the Negev – 2nd May 2010 So What Are Memetic Algorithms? MAs are carefully orchestrated interplay between (stochastic) global search and (stochastic) local search algorithms N. Krasnogor. Handbook of Natural Computation, chapter Memetic Algorithms. Natural Computing. Springer Berlin / Heidelberg, 2009 Ben Gurion University of the Negev – 2nd May 2010 • MAs are one of the key methodologies behind successful discrete/continuous optimisation – Nature Inspired Methodology? – A Research Paradigm • Key Design Issues underpinning MAs • A Pattern Language for MAs design Ben Gurion University of the Negev – 2nd May 2010 The Canonical MA At design From Eiben’s & Smith “Introduction To Evolutionary Computation” Ben Gurion University of the Negev – 2nd May 2010 time lots of issues arise Design Patterns and Pattern Languages In Alexander, C., Ishikawa, S., Silverstein, M., Jacobson, M., Fiksdahl-King, I., Angel, S.: A Pattern Language - Towns, Buildings, Construction. Oxford University Press (1977): “In this book, we present one possible pattern language,... The elements of this language are entities called patterns. Each pattern describes a problem which occurs over and over again in our environment, and then describes the core of the solution to that problem, in such a way that you can use this solution a million times over, without ever doing it the same way twice. “ Ben Gurion University of the Negev – 2nd May 2010 How are Patterns Described? High Level • Pattern name • Problem statement • The solution • The Consequences • Examples A collection of well deﬁned patterns, i.e. a rich pattern language, substantially enhances our ability to communicate solutions to recurring problems without the need to discuss speciﬁc implementation details. Ben Gurion University of the Negev – 2nd May 2010 Invariants and Decorations A Compact “Memetic” Algorithm by Merz (2003) Ben Gurion University of the Negev – 2nd May 2010 Invariants and Decorations A “Memetic” Particles Swarm Optimisation by Petalas et al (2007) Ben Gurion University of the Negev – 2nd May 2010 Invariants and Decorations A “Memetic” Artificial Immune System by Yanga et al (2008) Ben Gurion University of the Negev – 2nd May 2010 Invariants and Decorations A “Memetic” Learning Classifier System by Bacardit & Krasnogor (2009) Ben Gurion University of the Negev – 2nd May 2010 Invariants and Decorations • Many more based on Ant Colony Optimisation, NN, Tabu Search, SA, DE, etc. • Key Invariants: – Global search mode – Local search mode • Many Decorations, e.g.: – – – – Crossover/Mutations (EAs based MAs) Pheromones updates (ACO based MAs) Clonal selection/Hypermutations (AIS based MAs) etc Ben Gurion University of the Negev – 2nd May 2010 Memetic Algorithm Pattern (MAP) • Problem: how to successfully orchestrate local and global search • Solution: for a given domain find exploration and exploitation mechanisms that work in synergy. • Consequence: increase CPU? Diversity lost? • Examples: too many! Ben Gurion University of the Negev – 2nd May 2010 A Pattern Language for Memetic Algorithms Memetic Algorithms by N. Krasnogor. Handbook of Natural Computation (chapter) in Natural Computing. Springer Berlin / Heidelberg, 2009. www.cs.nott.ac.uk/~nxk/publications.html Ben Gurion University of the Negev – 2nd May 2010 Outline of the Talk – Evolutionary Algorithms Revisited – MAs’ Motivation (General Versus Problem Specific Solvers) – MAs design issues – A Few Exemplar Applications – Related Methods and Advanced Topics – Putting it all Together – Conclusions, Q&A Ben Gurion University of the Negev – 2nd May 2010 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 PhD 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. Ben Gurion University of the Negev – 2nd May 2010 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 • Consider whether Baldwinian or Lamarckianism is better •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. Ben Gurion University of the Negev – 2nd May 2010 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! Ben Gurion University of the Negev – 2nd May 2010 The Future of MAs Software Growth • Software should be “seeded” and grown, very much like a plant or animal (including humans) • Software should be embryonic and develop when situated on a production environment • What would a software “incubation” machine look like? • What would a software “germinator” look like? Ben Gurion University of the Negev – 2nd May 2010 Organs Individual Cells Tissue DNA/RNA Ben Gurion University of the Negev – 2nd May 2010 Production Environment Input TSP Organ SC SC Software Cell Euclidean TSP Organ SC SC SC SC Pluripotential Solver “DNA” GraphicalTSP Organ Ben Gurion University of the Negev – 2nd May 2010 TSP Solver Software Organism Questions?!? Thank you: Professor Dr. Moshe Sipper for inviting me to give this talk Ben Gurion University of the Negev – 2nd May 2010

Descargar
# Recent Advances in Memetic Algorithms