Josu Ceberio
Previously… EDAs for integer domains. EDAs for real value domains. Few efficient designs for permutation- based problems. POOR PERFORMANCE EHBSA and NHBSA (Tsutsui et al.)
Distance-based ranking models The Mallows model is a distance-based exponential model. Two parameters Consensus ranking, Spread parameter, Probability distribution
Distance-based ranking models Kendall’s tau distance Decomposition of the distance Factorization of the probability distribution
Distance-based ranking EDA Generalized Mallows EDA is proposed. A generalization of the Mallows model. spread parameters. Probability distribution
The problem To check the performance we approach: Permutation Flowshop Scheduling Problem. Extensively studied. The Mallows EDA demonstrated good performance.
Permutation Flowshop Scheduling Problem Given a set of n jobs and m machines and processing times p ij. Find the sequence for scheduling jobs optimally. Optimization criterion: Total Flow Time (TFT). Codification m1m1 m2m2 m3m3 m4m4 j1j1 j3j3 j2j2 j5j5 j4j4 Example Objective function
Generalized Mallows EDA Preliminary experiments Spread parameters
Generalized Mallows EDA Preliminary experiments GM model convergence
Generalized Mallows EDA Approximating spread parameters Newton-Raphson An upper bound for the spread parameters is fixed!!
Generalized Mallows EDA Approximating spread parameters
Standart evolutionary shape Restart mechanism shape Generalized Mallows EDA Preliminary experiments Restart mechanism Improvement !
PFSP state-of-the-art LR(n/m) GA VNS Crossover VNS Asynchronus Genetic Algorithm (AGA) – Xu et al Local Search (Swap) Local Search (Insert) Shake
PFSP state-of-the-art LR(n/m) Local Search (Swap) Local Search (Insert) Shake Variable Neighborhood Search 4 (VNS 4 ) – Costa et al. 2012
PFSP state-of-the-art Fundamentalist approaches rarely achieve optimum solutions. Hybridization is the path to follow. High presence of VNS algorithms.
First approach to the PFSP GM-EDA does not succeed. An hybrid approach is considered: Hybrid Generalized Mallows EDA (HGM-EDA)
Hybrid Generalized Mallows EDA Generalized Mallows EDA Local Search (Swap) Local Search (Insert) Orbit Shake VNS
Experimentation Algorithms: AGA, VNS 4, GM-EDA, VNS and HGM-EDA. 20 repetitions Taillard’s PFSP benchmarks: 100 instances 20 x x x x x x x x x x x x 20
Experimentation Spread parameters upper bound. Select the upper-theta that provides the best solutions for GM-EDA Stopping criterion: maximum number of evaluations. Evaluations performed by AGA in n x m x 0.4s.
Experimentation Taillards benchmark 20 x 520 x 1020 x 20 AGA VNS GM-EDA VNS HGM-EDA
Experimentation Taillards benchmark 50 x 550 x 1050 x 20 AGA VNS GM-EDA VNS HGM-EDA
Experimentation Taillards benchmark 100 x 5100 x x 20 AGA VNS GM-EDA VNS HGM-EDA
Experimentation Taillards benchmark 200 x x x 20 AGA VNS GM-EDA VNS HGM-EDA
Experimentation Taillard’s benchmark - Summary 20x0520x1020x2050x0550x1050x20100x05100x10100x20200x10200x20500x20 AGA ✔✔✔✔✔✔✔✔ VNS 4 ✔✔✔ GM-EDA VNS ✔✔✔ HGM-EDA ✔✔✔✔✔✔✔
Experimentation Taillard’s benchmark – Results analysis HGM-EDA outperforms state-of-the-art results in some cases. ○ Which is the reason for the performance fall given in instances of 500x20? Biased instances? -A tabu search algorithm was used for to choose the hardest instances. We generate a random benchmark
Experimentation Random benchmark New configurations between 200 and 500. Total: 100 instances. 250 x x x x x x x x x x 20
Experimentation Random benchmark - Summary 250x10250x20300x10300x20350x x20400x10400x20450x10450x20 AGA ✔✔✔ VNS 4 GM-EDA VNS HGM-EDA ✔✔✔✔✔✔✔
Experimentation Random benchmark – Results analysis Statistical Analysis confirms experimentation. ○ Friedman test + Shaffer’s static. HGM-EDA and AGA are definitely the best algorithms. VNS 4 results do not match with those reported. The performance falls onwards 400x20. What’s wrong with largest instances?
Analysis – Hybrid approach Improvement ratio EDA vs. VNS
Analysis – Generalized Mallows EDA AGA vs. GM-EDA
Analysis – Generalized Mallows EDA Thetas convergence
Stops prematurely!!!
Analysis – HGM-EDA vs. AGA More evaluations Max eval.AGAHGM-EDA x x x x x x x x One instance of 500x20
Analysis – Generalized Mallows EDA LR vs. GM-EDA
Analysis – HGM-EDA vs. AGA More evaluations Max eval.AGAHGM-EDA x x x x x x x x One instance of 500x20
Analysis – HGM-EDA vs. AGA More evaluations Max eval.AGAHGM-EDAGuided HGM-EDA x x x x x x x x One instance of 500x20
Analysis – HGM-EDA vs. AGA More evaluations One instance of 500x20
Conclusions Hybrid Generalized Mallows EDA is a efficient algorithm for solving the PFSP. Succeed in 152/220 instances. The participation of the GM-EDA is essential.
Future Work - PFSP Test other parameters: evaluations, population size, theta bounds, selection size… Include information of the instance. Guided Initialization Shake the solution of the LR(n/m) to build up the population?
Future Work – GM-EDA Set different upper bounds to the spread parameters Study other distances. Is suitable Kendall’s-tau distance? Other distances: Cayley, Ulam, Hamming Study the problem. Other problems: TSP QAP LOP (work in progress)
Eskerrik asko Josu Ceberio Eskerrik asko Josu Ceberio
Distance-based ranking EDA Mallows EDA Learning and Sampling 0...n n - 1