Learning Guided Multiobjective Optimization Aimin Zhou East China Normal University, Shanghai, China 7/9, 2015 1 Outline o Evolutionary Multiobjective Optimization o A Self-Organizing Map based Approach o Learning Guided Evolution – A Short Survey o Conclusions & Future Remarks 2 LGMO - A.Zhou @ ECNU 7/9,2015 Outline o Evolutionary Multiobjective Optimization o A Self-Organizing Map based Approach o Learning Guided Evolution – A Short Survey o Conclusions & Future Remarks 3 LGMO - A.Zhou @ ECNU 7/9,2015 Multiobjective Optimization Problem o MOP min F(x) = ( f1 (x), f 2 (x),… , f m (x)) s.t x Î D where D : feasible region of decision variables. fi : D ® R, objective function F : D ® R m , objective vector function F(D) = {F(x) | x Î D} : attainable objective set o real-world applications o scientific and engineering problems 4 D x F f2 F (D ) z z ( z1,z2 ) F ( x ) f1 z1 f1 ( x ) , z2 f 2 ( x ) LGMO - A.Zhou @ ECNU 7/9,2015 Optimum of an MOP o For a minimization problem Let x, y Î D, x dominates y (or F(x) dominates F(y)) f i (x) £ f i (y) for all i and f j (x) < f j (y) for at least one index j. o dominate = be better than o Examples: x (z ) dominates x (z ). 3 3 1 1 (x ) z and x (z ) cannot be compared with each other. 2 2 1 x1 x2 x3 D F f2 z2 z1 1 F (D ) z3 f1 domination is a partial ordering why MOPs are harder than single opt. problems 5 LGMO - A.Zhou @ ECNU 7/9,2015 Optimum of an MOP o Pareto optimal solution Pareto set (PS) a solution cannot be dominated by any other solutions. o Pareto set (PS) the set of all the Pareto optimal solutions in decision variable space. f2 F o Pareto front (PF) PF=F(PS) (in objective space) F (D ) Pareto front (PF) f1 The PF is the southwest boundary of F(D). 6 LGMO - A.Zhou @ ECNU 7/9,2015 Task of MOEA Very often, a decision maker wants Pareto set (P) A representative set of Pareto optimal solutions (uniformly distributed along the PF or PS) f2 F F (D ) Task of most Multiobjective Evolutionary Algorithms (MOEAs) Pareto front (PF) f1 [1] A. Zhou, B. Qu, H. Li, S. Zhao, P. Suganthan, and Q. Zhang, Multiobjective evolutionary algorithms: A survey of the state of the art, Swarm and Evolutionary Computation, 1(1): 32–49, 2011. 7 LGMO - A.Zhou @ ECNU 7/9,2015 Outline o Evolutionary Multiobjective Optimization o A Self-Organizing Map based Approach o Learning Guided Evolution – A Short Survey o Conclusions & Future Remarks 8 LGMO - A.Zhou @ ECNU 7/9,2015 Motivation o Regularity of continuous MOPs: Pareto set (PS) Under certain conditions, the PS (PF) is a (m-1)-dimensional piecewise continuous manifold in decision (objective) space. (m is the # of the objs.) f2 F o Problem-specific knowledge is useful for algorithm design. How can we deal with a continuous MOP if its PS is (m-1)-D piecewise continuous manifold? F (D ) Pareto front (PF) f1 [1] Q. Zhang, A. Zhou, and Y. Jin, RM-MEDA: a regularity model-based multiobjective estimation of distribution algorithm, IEEE Transactions on Evolutionary Computation, 12(1):797-799, 2008. 9 LGMO - A.Zhou @ ECNU 7/9,2015 Motivation o Classical reproduction operators scalar-objective optimization x2 x2 x2 x* x2 b A A a B a x1 B x1 (b) 单点杂交 (d) 高斯模型采样 x2 x2 PS x2 PS a B A b x1 PS a b PS B A x1 (b) 单点杂交 x1 (c) 算术杂交 multiobjective optimization (a) 当前种群 b x1 (a) 当前种群 x2 x* x* x* x1 (c) 算术杂交 x1 (d) 高斯模型采样 [1] A. Zhou, Q. Zhang, and G. Zhang, Multiobjective evolutionary algorithm based on mixture Gaussian models, Journal of Software, 25(5):913-928, 2014. 10 LGMO - A.Zhou @ ECNU 7/9,2015 Basic Idea o Algorithm framework Population Reproduction operators Competition Replacement Selection (Replacement): quite a lot of works Reproduction: our focus New Solutions 11 LGMO - A.Zhou @ ECNU 7/9,2015 Self-Organizing Maps Z1 D1 X3 R3 Neuron node Latent space u D2 O wun 1 i Z2 n x=(x1 , ... ,xi, ... ,xn) Iutput space X2 pu Q X1 o MOP o SOM regularity property latent model mating registration similarity detection [1] H. Zhang, A. Zhou, S. Song, Q. Zhang, X. Gao, and J. Zhang, A self-organizing multiobjective evolutionary algorithm, 2015 (submit). 12 LGMO - A.Zhou @ ECNU 7/9,2015 SOM Assisted MOEA o Characteristics: Call SOM and MOEA main steps iteratively detect the population structure in an incremental manner save computational cost Generate offspring by neighboring parents 13 LGMO - A.Zhou @ ECNU 7/9,2015 Other Issues o Reproduction operator: Differential Evolution (DE) Polynominal Mutation 14 o Selection operator: Nondominated sorting scheme LGMO - A.Zhou @ ECNU 7/9,2015 Experimental Results o On irregular problems GLT test suite CellDE, MOEA/D-DE, RM-MEDA, NSGA-II, SMS-EMOA,SOM-NSGA-II IGD,HV metrics 15 LGMO - A.Zhou @ ECNU 7/9,2015 Experimental Results o Run time performance Converges faster in most cases. 16 LGMO - A.Zhou @ ECNU 7/9,2015 Experimental Results o Visual performance 17 LGMO - A.Zhou @ ECNU 7/9,2015 Experimental Results o Visual performance 18 LGMO - A.Zhou @ ECNU 7/9,2015 Outline o Evolutionary Multiobjective Optimization o A Self-Organizing Map based Approach o Learning Guided Evolution – A Short Survey o Conclusions & Future Remarks 19 LGMO - A.Zhou @ ECNU 7/9,2015 Basic Questions Learning + Evolutionary Optimization o What? Learning Guided Evolution (LGE) is a kind of evolutionary algorithms that utilize statistical and machine learning techniques to guide the search. o Why? Priori & learnt problem specific knowledge to guide the search, and thus to improve search performance. o How? data organization pattern recognition pattern usage 20 initialization reproduction selection stop condition LGMO - A.Zhou @ ECNU 7/9,2015 Related Work o Adaptive Evolution Parameter tuning Operator selection mine populations Stopping condition o Estimation of Distribution Algorithm (EDA) Ant Colony Optimization (ACO) Cross-entropy method (CE) model & sample populations Covariance Matrix Adaptation Evolution Strategy (CMA-ES) o Surrogate Assist Evolutionary Algorithm (SAEA) replace evaluation 21 LGMO - A.Zhou @ ECNU 7/9,2015 Taxonomy o Angle of Machine Learning Regression based EAs Supervised Evolution Classification based EAs Manifold learning based EAs Learning Guided Evolution Unsupervised Evolution Clustering based EAs Density estimation based EAs Semisupervised Evolution 22 LGMO - A.Zhou @ ECNU 7/9,2015 A Short Survey of Our Recent Work o Regression based approaches Surrogate assisted minimax optimization Time series prediction for dynamic multiobjective optimization Cheap surrogate model PS estimation = PS manifold learning + center point prediction [1] A. Zhou, and Q. Zhang, A surrogate-assisted evolutionary algorithm for minimax optimization, in IEEE Congress on Evolutionary Computation (CEC 2010), Barcelona: IEEE Press, 2010, pp.1-7. [2] A. Zhou, Y. Jin, and Q. Zhang, A population prediction strategy for evolutionary dynamic multiobjective optimization, IEEE Transactions on Cybernetics, 44(1):40-53,2014. [3] A. Zhou, J. Sun, and Q. Zhang, An estimation of distribution algorithm with cheap and expensive local search, IEEE Transactions on Evolutionary Computation, 2015. (accepted) 23 LGMO - A.Zhou @ ECNU 7/9,2015 A Short Survey of Our Recent Work o Classification based approaches Classification based preselection Classification based selection selection = classification [1] J. Zhang, A. Zhou, and G. Zhang, A Classification and Pareto domination based multiobjective evolutionary algorithm, in Proceedings of IEEE Congress on Evolutionary Computation (CEC 2015), 2015, pp.1-8. [2] J. Zhang, A. Zhou, and G. Zhang, A classification based preselection for evolutionary algorithms, 2015 (submit). 24 LGMO - A.Zhou @ ECNU 7/9,2015 A Short Survey of Our Recent Work o Manifold learning based approaches Regularity model based multiobjective estimation of distribution algorithm (RM-MEDA) 1 population 2 C1 3 x1 C2 x2 C3 x [1] Q. Zhang, A. Zhou, and Y. Jin, RM-MEDA: a regularity model-based multiobjective estimation of distribution algorithm, IEEE Transactions on Evolutionary Computation, 12(1):797-799, 2008. [2] A. Zhou, Q. Zhang, and Y. Jin, Approximating the set of Pareto-optimal solutions in both the decision and objective spaces by an estimation of distribution algorithm, IEEE Transactions on Evolutionary Computation, 13(5):1167-1189, 2009. 25 LGMO - A.Zhou @ ECNU 7/9,2015 A Short Survey of Our Recent Work o Clustering based approaches Clustering based mating selection Self-organizing multiobjective evolutionary algorithm [1] H. Zhang, S. Song, and A. Zhou, A clustering based multiobjective evolutionary algorithm, in IEEE Congress on Evolutionary Computation (CEC 2014), 2014. [2] H. Zhang, A. Zhou, S. Song, X. Gao, and J. Zhang, A self-organising multiobjective evolutionary algorithm, 2015. (submit) 26 LGMO - A.Zhou @ ECNU 7/9,2015 A Short Survey of Our Recent Work o Density estimation based approaches Mixture Gaussian model model base reproduction model re-use Non-parametric density estimation model based pre-selection multi-operator search locally weighted model fitness estimation by cheap models [1] L. Zhou, A. Zhou, G. Zhang, C. Shi, An estimation of distribution algorithm based on nonparametric density estimation, in IEEE Congress on Evolutionary Computation (CEC 2011), New Orleans: IEEE Press, 2011, pp.1597-1604. [2] A. Zhou, Q. Zhang, and G. Zhang, A multiobjective evolutionary algorithm based on decomposition and probability model, in IEEE Congress of Evolutionary Computation (CEC 2012), Brisbane: IEEE Press, 2012, pp.1-8. [3] A. Zhou, Q. Zhang, and G. Zhang, A multiobjective evolutionary algorithm based on mixture Gaussian models, Journal of Software, 25(5):913−928, 2014. [4] Q. Liao, A. Zhou, and G. Zhang, A locally weighted metamodel for pre-selection in evolutionary optimization, in The IEEE Congress on Evolutionary Computation (CEC 2014), 2014. [5] A. Zhou, Y. Zhang, G. Zhang, and W. Gong, On neighborhood exploration and subproblem exploitation in decomposition based multiobjective evolutionary algorithms, in Proceedings of IEEE Congress on Evolutionary Computation (CEC 2015), 2015, pp.1-8. [6] W. Gong, A. Zhou, and Z. Cai, A multi-operator search strategy based on cheap surrogate models for evolutionary optimization, IEEE Transactions on Evolutionary Computation, 2015. (accepted) 27 LGMO - A.Zhou @ ECNU 7/9,2015 A Short Survey of Our Recent Work o Adaptive approaches Adaptive replacement strategy in MOEA/D cost Adaptive resource allocation in MOEA/D subproblem index resource control [1] Z. Wang, Q. Zhang, A. Zhou, M. Gong, and L. Jiao, Adaptive replacement strategies for MOEA/D, IEEE Transactions on Cybernetics, 2015. (accepted) [2] A. Zhou, and Q. Zhang, Are all the subproblems equally important? Resource allocation in decomposition based multiobjective evolutionary algorithms, IEEE Transactions on Evolutionary Computation, 2015. (accepted) 28 LGMO - A.Zhou @ ECNU 7/9,2015 Outline o Evolutionary Multiobjective Optimization o A Self-Organizing Map based Approach o Learning Guided Evolution – A Short Survey o Conclusions & Future Remarks 29 LGMO - A.Zhou @ ECNU 7/9,2015 Conclusions & Future Remarks Cost Alg. Cost Problem Cost o Random Search: Alg. Cost is LOW, Problem Cost is HIGH. o Mathematical Programming: Alg. Cost is HIGH, Problem Cost is LOW. o Evolutionary Optimization: BETWEEN the above two approaches. o Learning Guided Evolutionary Optimization o It Is promising to balance the two costs. o There is no systematic study yet. o Which knowledge to detect? o Which learning method to use? o How to combine learning methods and evolutionary algorithms? 30 LGMO - A.Zhou @ ECNU 7/9,2015 Thanks! Dr. Aimin Zhou, East China Normal University amzhou@cs.ecnu.edu.cn, http://www.cs.ecnu.edu.cn/~amzhou http://faculty.ecnu.edu.cn/s/1949/t/22630/main.jspy 31 LGMO - A.Zhou @ ECNU 7/9,2015