Omni-Optimizer A Procedure for Single and Multi-objective Optimization Prof. Kalyanmoy Deb and Santosh Tiwari
Motivation Generic Programming Practices Unified algorithm for all types of optimization problems An efficient implementation of NSGA-II framework (procedure) Developing an efficient and self-adaptive optimization paradigm
Literature Survey CHC (Cross generation elitist selection, Heterogeneous recombination, Cataclysmic mutation) – Explicit Diversity GENITOR (Steady state GA), more like (µ+1)-ES so far as selection mechanism is concerned. – High selection pressure NPGA (Niched Pareto Genetic Algorithm), uses sharing parameter σ share – # of niches obtained depend on the sharing parameter
Literature Survey contd… PESA (Pareto Envelope-based Selection Algorithm), Hyper- grid division of phenotypic space, selection based on crowding measure NSGA-II (Non-dominated Sorting Genetic Algorithm) SPEA2 (Strength Pareto Evolutionary Algorithm), Fine grained fitness assignment mechanism utilizing density information, Only archive members participate in mating – Excellent Diversity in phenotypic space NCGA (Neighborhood Cultivation Genetic Algorithm), used neighborhood crossover, based on NSGA-II and SPEA2 RPSGAe (Reduced Pareto Set Genetic Algorithm with elitism) ENORA (Evolutionary Algorithm of Non-dominated Sorting with Radial Slots)
Salient Features of the Algorithm Based on NSGA-II framework Based on the concept of Pareto dominance Incorporates elitism Explicit diversity preserving mechanism Can be used for single-objective as well as multi- objective problems Can be used for uni-global as well as multi- global problems Independent of the number of niches that an optimization problems exhibits
Moving beyond NSGA-II Restricted Selection Scheme Tournament selection based on usual domination Non-dominated sorting based on epsilon dominance Crowding Distance Assignment Genotypic as well as Phenotypic space niching Choose best members from above average population Remove worst members from below average population More robust recombination and variation operators Two point crossover for binary variables Highly disruptive real variable mutation
Restricted Selection Helps in preserving multi-modality Experiments show that it gives faster overall convergence
Epsilon Domination Principle A finite percentage (based on function value) of solutions assigned a particular rank A finite percentage (based on function value) of solutions assigned a particular rank Allows somewhat inferior solutions to remain in the population Allows somewhat inferior solutions to remain in the population Provides guaranteed diversity Provides guaranteed diversity Helps to obtain multi-modal solutions in case of single objective problems Helps to obtain multi-modal solutions in case of single objective problems Epsilon is generally taken in the range ~ Epsilon is generally taken in the range ~ 10 -6
Modified Crowding Distance Genotypic as well as Phenotypic space niching
Highly Disruptive Mutation Operator
Computational Complexity Restricted selection O (nN 2 ) Ranking procedure O (MN 2 ) Crowding procedure max{ O (MN log N), O (nN log N) } Overall iteration-wise complexity max {O (nN 2 ), O (MN 2 ), O (nN log N)}
Implementation Details Code written in simple C and strictly conforms to ANSI/ISO standard Data structure used is a custom doubly linked list (gives O(1) insertion and deletion) Randomized quick sort used for sorting Almost all the functions perform in-place operation (addresses are passed, significantly decreases stack overheads)
Simulation Results GA parameters for all problems chosen as follows η c = 20 η m = 20 P (crossover) = 0.8 P (mutation) = 1/n, where n = # of real variables δ = Population size and number of generations different for different problems
Simulation Results contd… 20 variable Rastrigin function # of function evaluation Least = Median = Worst = 20 variable Schwefel function # of function evaluation Least = Median = Worst = Other single objective problems can be found in the paper In all cases, better results are found in comparison to previous reported studies
Single objective multi-modal function f(x) = sin 2 (πx)x є [0,20]
Single objective multi-modal function Unconstrained Himmelblau’s function
Multi-objective Uni-Global Test Problems 30 variable ZDT2 (100×100)
Multi-objective Uni-Global Test Problems 10 variable ZDT4 (100×250)
Multi-objective Uni-Global Test Problems CTP4 (100×7000)
Multi-objective Uni-Global Test Problems CTP8 (100×100)
Multi-objective Uni-Global Test Problems DTLZ4 (300×100)
Multi-objective Multi-Global Test Problem F 1 (x) = summation (sin (πx i ) )x i є [0,6] F 2 (x) = summation (cos (πx i ) )x i є [0,6] Efficient points in phenotypic space
Multi-objective Multi-Global Test Problem Genotypic space plots
Few Sample Simulations F(x) = sin 2 (10,000*pi*x) Himmelblau’s Functions ZDT Test Problems CTP Test Problems Test Problem TNK Multi-global Multi-objective Test Problem
Further Ideas and Future Work Incorporating PCX instead of SBX for crossover Automatically fine-tuning mutation index so as to achieve arbitrary precision Self-adaptation of parameter δ Segregating population into niches without the introduction of DM Dynamic population sizing Using hierarchical NDS for the crowding distance assignment