Evolutionary of the Variable-Length Multi-objective Genetic Algorithm 李宗南 國立中山大學 資訊工程學系 December 3, 2008 2 Outline • A tale of single objective optimization and multi-objective optimization • The single genetic algorithm • The multi-objective genetic algorithm • The variable length genetic algorithm • The variable length multi-objective optimization • Applications - Aircraft routing - Placement of heterogeneous wireless transmitters • Conclusions 3 A Tale of Single Objective Optimization and Multi-Objective Optimization Courtesy of Dr. YoaChu Jin Single Objective Optimization Given: a function f : A → R from some set A to the real numbers Sought: an element x0 in A such that f(x0) ≤ f(x) for all x in A ("minimization") or such that f(x0) ≥ f(x) for all x in A ("maximization"). 5 Algorithms for Single Objective Optimization Gradient descent aka steepest descent or steepest ascent Hill climbing Simulated annealing Quantum annealing Tabu search Beam search Genetic algorithms Ant colony optimization Evolution strategy Stochastic tunneling Differential evolution Particle swarm optimization Harmony search Bees algorithm Dynamic relaxation 6 Multi-Objective Optimization Courtesy of Dr. YoaChu Jin 7 Multi-Objective Optimization m= 1 , single objective optimization Courtesy of Dr. YoaChu Jin 8 8 Solutions for Multi-objective Optimization • Map into a single objective optimization by a weighted sum • The multi-objective approach (rank-based fitness assignment method) to evaluate each objective individually 9 Comparison of the single objective approach and the multi-objective approach SO + simple - Hard to determine weight for each objective . - Hard to prevent some objectives from dominating others. MO + Have the ideal situation where each objective function attains a satisfactory level. + Have the flexibility to achieve different levels of tradeoff. - Not so easy to solve. 10 Single GA 11 Single GA Parent A 110011001 Parent B 101111011 110011001 …. …. 101010101 110011001 101111011 …. 101010101 110011001 => 110101001 110011001 101111011 110011011 101111001 110011001 101111011 …. 101010101 12 Introduction of multi-objective genetic algorithm Task 1: To find a set of solutions as close as possible to the Pareto-optimal front. Task 2: To find a set of solutions as diverse as possible f2 Dominated solutions Nondominated solutions Solution Space Pareto Front Diversity proximity Minimization Objective 1 f1 13 The variable length genetic algorithm Why Variable-Length GAs (VLGAs)? The number of solutions is not fixed. i.e. Fixed Length GAs must know number of variables a priori Ex: Finding number of base stations for a given region Ex: Finding rules for autonomous agents 14 Evolutionary of the variable length multiobjective optimization The number of solutions is not fixed. It is a multi-objective optimization problem We would like to solve the problem by GA 15 Evolutionary of the variable length multiobjective optimization Nondominated solutions f2 Solution Space Dominated solutions Nondominated solution Pareto Front Diversity proximity Minimization Objective 1 f1 16 Evolutionary of the variable length multiobjective optimization Rank-based fitness assignment method f2 3 3 5 1 8 3 Front 4 1 5 Front 3 2 4 1 1 1 Front 2 Front 1 1 f1 17 Evolutionary of the variable length multiobjective optimization 18 Applications MOGA - Aircraft routing VLMOGA - Placement of heterogeneous wireless transmitters 19 Application 1 Aircraft Routing using Multiobjective Genetic Algorithm 20 Problem Description • Aircraft routing ▫ A given set of flights a group of aircrafts ▫ Available amount of aircrafts 21 Aircraft Routing (1/2) Timetable assign to aircrafts Flight ID Dep. Time Arv. Time Origin Destination f1 f2 f3 f4 f5 f6 f7 f8 09:15 19:00 14:00 11:20 22:00 16:30 08:10 20:35 09:50 19:35 14:35 11:55 22:50 17:20 09:00 21:25 KHH TSA MZG TSA TNN TSA KHH KHH TSA KHH TSA MZG TSA TNN TSA TSA f9 18:00 18:50 TNN TSA f10 15:30 16:20 TSA TNN Flight ID Dep. Time Arv. Time f11 f12 f13 f14 f15 f16 f17 f18 f19 f20 09:30 19:10 14:10 10:20 18:00 15:30 08:20 18:35 13:00 10:20 10:05 19:45 14:45 10:55 18:50 16:20 09:20 19:25 13:35 10:55 Origin Destination HUN TSA MZG TSA TXG TSA KHH KHH MZG TSA TSA HUN TSA MZG TSA TXG TSA TSA TSA MZG 22 Aircraft Routing (2/2) f2 f1 f3 f4 f12 f14 f11 f8 f9 f6 f17 f5 f15 f7 f16 f18 f13 f10 f19 f20 Flight set F aircraft 1 f8 f2 f13 f11 aircraft 2 f3 f6 f4 f10 f17 aircraft 3 f18 f16 f19 f9 f5 aircraft 4 f20 f1 f14 f7 f12 f15 Flight schedule S 23 Notations • Let α, β, ω, and γ represent ▫ ▫ ▫ ▫ α: β: ω: γ: number of aircrafts maximal number of flights per aircraft number of airports number of daily flights, respectively. • Set of flights: F = {fi|1 ≤ i ≤ γ} • Set of airports: P = {pj|1 ≤j ≤ ω} P= {台北松山機場、 高雄小港、 台中、 台南 、馬公、 金門} 24 , Associate Information of One Flight The flight schedule S can be represented as: si,j: the jth flight assigned to the ith aircraft : origin of si,j, where : destination of si,j, where : departure time from : arrival time in 25 Definition of a Flight Schedule Maximal flights assigned to each aircraft Number of aircrafts Flight Schedule S Number of aircrafts Maximal flights assigned to each aircraft s1,1 sα,β 2 6 27 Objectives Objectives: Subject to • Ground turn-around time objective • Flow balance objective 28 Ground Turn-around Objective(1) Legal ground turn-around time: TGH Taipei P Kaohsiung Q Makung Δt 29 Flow Balance Objective(2) Taipei Taipei f1 f1 Kaohsiung Kaosiung f2 Makung f2 Makung (a) (b) Extra cost 30 Encoding Scheme 1 2 … β-1 β c1 s1,1 s1, 2 … s1, β-1 s1, β c2 s2,1 s2, 2 … s2, β-1 s2, β : : : : : : cα sα, 1 sα, 2 … sα, β-1 sα, β s1,1 s1, 2 … s1, β-1 flights of c1 s1, β s2,1 s2, 2 … s2, β-1 flights of c2 s2, β … … … sα, 1 sα, 2 … sα, β-1 flights of cα sα, β 31 Crossover A B C D E F Mapping exchange relationship B C E A F D cutpoint A B C A F D mapping relationship E A A B C E F D duplicate genes B C E D E F 1. change A to E B C E D A F 2. change E to A 32 Reciprocal Mutation A B C D E F A B F E D C Experimental Results • 7 airports, 9 aircrafts, 12 flights one day, 79 flights. Airports HUN KHH KNH MZG TNN TSA Parameter Value Crossover rate 1 Mutation rate 0.2 No. of generations 5000 Population size 100 Reproduction rate 0.8 TTT 33 34 Experimental Results 35 Symbols in Experimental Results Gantt chart: time Aircraft1 /crew1 Aircraft2 /crew2 departure time origin flight ID 682 KNH TSA destination arrival time 36 Scheduling Result of 9 aircrafts 37 Scheduling Result of 8 aircrafts Example ε = [ε1, ε2] =[k1 × α × TGH, k2× α] = [1 × 8 × 25, 1 × 8] = [200, 8] Original value Φ(S) = [ϕ1(S), ϕ2(S)] Values in auxiliary vector of performance indices Λ(S, ε)=[λ(S,ε1), λ(S,ε2)] Result 1 [195,0] [0,0] Result 2 [190,1] [0,0] Result 3 [185,2] [0,0] 38 Scheduling Result of 8 aircrafts Result1 Result2 Result3 39 Scheduling Result of 8 Aircrafts 40 Retiming Process Retiming process Station K Station K P P Station K+1 Station K+1 Q Q Station K+2 Station K+2 Inspection line Flights P and Q cannot be assigned to the same aircraft Inspection line Flights P and Q can be assigned to the same aircraft 41 Scheduling Result-Retiming 42 Application 2: Heterogeneous Wireless Transmitter Placement with Multiple Constraints based on the Variable-Length Multi-objective Genetic Algorithm 43 Problem statement Choose a set of heterogeneous wireless transmitters to place on the designed space to fulfill certain design requirements such as Position, power, capacity, frequency channel assignment, overlap, data rate demand, population density, cost and coverage Evolutionary multiobjective optimization for base station transmitter placement with frequency assignment, IEEE Trans. on Evolutionary Computation, 2003 44 Introduction (cont.) 23meters 15meters 45 Parameters Parameters Type of transmitters Value 1 Transmitter cost Maximum allowed power loss threshold Transmitter capacity 2200 63.5dB (radius 15 meters) Test points Total 10404 points Data Rate Demand 21kbps Generation 3000 54Mb Wireless transmitter placement problem 47 Problem Definition • Model ▫ Map, receiver, transmitter • Receiver ▫ Position, data rate demand, sensitivity • Transmitter ▫ Position, type=(cost, power, capacity) 48 Path Loss Propagation Models 1. Free space path loss model L(r i , t j ) 20 log 10( 4d ) 2. Log-distance path loss model with shadowing effect o ( r ,t ) L(r i , t j ) 20 log 10( 3. ECC-33 model 4d ) i j P (r , t ) A (r , t ) g 1 g i j g i j 49 Objectives • • • • Coverage Cost Data Rate Demand Overlap 50 Coverage Coverage Rate=4/9=44.4% Uncoverage=5 n Uncoverage (T ) cT (ri ), { i 1 cT (ri ) 1, if coverage T ( ri ) -1 cT (ri ) 0, otherwise 51 Cost • Cost=1000+300*2 =1600 Cost=1000 Cost=300 Cost=300 Cost (T ) cost (t j ) t j T 52 Data Rate Demand • Yellow: 16kbps • Blue: 128 kbps • Demand(t1)=128* 4+16*2=544 • Capacity(t1)=1024 • DRD(T)=|5441024|=480 DRD (T ) | demand ri t j T ri R , power ( t j ) L ( ri ,t j ) sensitivityi capacity (t j ) | Next generation wireless LAN system design, Proceedings on MILCOM, 2002 53 Overlap • Overlap=2 • Overlap rate=2/9 =22.2% Overlap (T ) overlap ri overlapri 0, if | AS ri | 0,1 ,{ overlapri 1, if | AS ri | 1 ri PG Evolutionary multiobjective optimization for base station transmitter placement with frequency assignment, IEEE Trans. on Evolutionary Computation, 2003 54 Objectives • Minimize n Uncoverage (T ) cT (ri ), { i 1 cT (ri ) 1, if coverageT ( ri ) -1 cT (ri ) 0, otherwise Cost (T ) cost (t j ) t j T DRD (T ) | demand ri t j T ri R , power ( t j ) L ( ri ,t j ) sensitivityi Overlap (T ) overlap ri PG ri capacity (t j ) | overlapri 0, if | AS ri | 0,1 ,{ overlapri 1, if | AS ri | 1 • Subject to ri , positioni PG, where 1 i n, n is the number of receivers t j , position j PG, where 1 j m, m is the number of transmitt ers 55 Encoding: Individual and Chromosome Representation An individual (x1,y1,z1),type1 (x2,y2,z2),type2 …… (xm,ym,zm),typem A chromosome X1 0 1 Y1 1 0 1 type1 Z1 0 1 0 Transmitter resolution= Encoding bits=3 0 0 56 Population Initialization • Temporary upper bound UB=10 • Random number=4 (1~10) type1 type1 type1 type1 type3 type3 type3 type2 type2 type1 Flowchart 57 Types of crossover • Uniform crossover for chromosome ▫ Change the position and type of transmitter • Variable-length one-point crossover ▫ Change the length of individual 58 Uniform Crossover for chromosome Individual1 (x1,y1,z1),type1 (x2,y2,z2),type2 Individual2 (x1,y1,z1),type1 (x2,y2,z2),type2 (x3,y3,z3),type3 59 One-point Crossover Split point Individual1 (x1,y1,z1),type1 (x2,y2,z2),type2 Individual2 (x1,y1,z1),type1 (x2,y2,z2),type2 Split point (x3,y3,z3),type3 60 Overall Crossover Parents Crossover Rate: Pc 1-Pc Pc Variable-length one-point crossover Uniform crossover for chromosome 1-Pc Pc Uniform crossover for chromosome 1 Copy from parents Offspring Flowchart 61 Mutation X1 chromosome1 0 1 Y1 1 0 X1 chromosome1 0 1 1 0 1 Y1 1 0 0 type1 Z1 0 0 type1 Z1 0 1 0 1 Pm length of chromosome 0 0 0 Flowchart 62 Types of Simulation Population Size Types of Transmitters Free space 100 1 Path loss 100 1 2D path loss 500 2 3D path loss 200 2 Upper bound UB=8,12,15 500 2 Heterogeneous 500 2 Indoor Outdoor 63 Overall Parameters Parameters Value Termination Crossover rate 5000 0.8 Mutation rate Pm 1/length of chromosome Frequency 2.4GHz 64 Indoor Freespace Parameters Parameters Penetration loss Type of transmitters Value Zero 1 Maximum allowed power loss threshold Transmitter capacity 66dB (radius 20 meters) Test points Every 3m x 3m, total 231 points 1024kbps Data Rate Demand 54Mb 65 Indoor Map • Floor plan of IC factory 66 Indoor Free Space • Threshold 66dB (radius 20 meters) Uncoverage 0 Cost 6600 Data rate 83968 demand Overlap 13 BS # 3 67 Indoor Path Loss • Type 1 –cement wall: 3.3dB • Type 2 – thickened cement wall: 6.5dB Uncoverage 0 Cost 13200 Data rate demand 30720 Overlap 48 BS # 6 68 Outdoor Map 69 Two Dimensional Outdoor Path Loss Parameters Parameters Penetration loss Types of transmitters Maximum allowed power loss threshold Transmitter cost Transmitters capacity Test points Value Type 1 –concrete wall: 8dB Type 2–mountain: 99dB 2 Type 1 – 103dB (1.5KM radius) Type 2 – 80dB (100 meter radius) Type 1 – 40000 Type 2 – 2200 Type 1 – 75Mb Type 2 – 54Mb Total 3057 points 70 Outdoor 2D-Data Rate Demand • Blue: 16 kbps/ 81% • Green:128kbps/ 13.7% • Red:1024kbps/ 5.2% 71 Outdoor 2D Result Result-11 Transmitters Uncoverage 25 Cost 480000 Data rate 403376 demand Overlap 2000 BS 1 # 12 BS 2 # 0 72 Solutions Idv 1 Idv 2 Idv 3 Idv 4 Idv 5 Idv 6 Idv 7 Idv 8 Uncoverage 36 67 1877 2106 1162 875 477 2941 Cost 440000 400000 40000 26400 80000 93200 13100 2200 Data rate 507776 399600 180352 847312 197648 515168 490832 277568 Overlap 1807 1114 0 79 35 16 248 0 BS 1 # 11 10 1 0 2 2 3 0 BS 2 # 0 0 0 12 0 6 5 1 73 Three Dimensional Outdoor Path Loss Parameters Parameters Types of transmitters Value 2 Maximum allowed power loss threshold Type 1 – 130dB (500 meter radius) Type 2 – 115dB (100 meter radius) Type 1 – 8000 Type 2 – 440 Type 1 – 75Mb Type 2 – 54Mb Total 5690 points Transmitter cost Transmitters capacity Test points 74 Outdoor 3D- Data Rate Demand • Blue: 16 kbps/ 57% • Green:128kbps/ 28% • Red:512kbps/ 15% 75 Outdoor 3D Result Uncoverage 0 Cost 32000 Data rate 1493184 demand Overlap 3516 BS 1 # 4 BS 2 # 0 76 Conclusions • Have introduced an evolutionary of the variable-length multi-objective genetic algorithm • Have presented the applications of MOGA and VLMOGA - Flight scheduling - The multiple constraints heterogeneous wireless transmitter placement 77 本投影片研究主要參與者 • • • • • • 周大源博士 中山大學 資工系 劉東官教授 高雄第一科大 丁川康教授 中正大學 資工系 吳建興 張慧君 黃振愷 78 References 1. 2. 3. 4. Chuan-Kang Ting, Chung-Nan Lee, Hui-Jin Chang, and Jain-Shing Wu “Wireless Heterogeneous Transmitter Placement Using Multi-Objective Variable-Length Genetic Algorithm” accepted to appear in IEEE Trans. on SMC, Part B Ta-Yuan Chow, T. K. Liu and Chung-Nan Lee, Chi-Ruey Jeng “Method of Inequality-Based Multiobjective Genetic Algorithm for Domestic Daily Aircraft Routing ", IEEE Trans. on SMC, Part A. Volume: 38, Issue: 2 . March 2008 Sibel Yaman and Chin-Hui Lee“ A Multi-Objective Programming Approach to Compromising Classification Performance Metrics”, IEEE International Workshop on Machine Learning for Signal Processing August 27, 2007 Yaochu Jin, “Evolutionary Multi-Objective Optimization”, Honda Research Institute Europe 79 Question & Suggestion 80 • Thank you for your attentions