Combinatorial Optimization Chapter 8, Essentials of Metaheuristics, 2013 Spring, 2014 Metaheuristics Byung-Hyun Ha R2 Outline Introduction Greedy Randomized Adaptive Search Procedures (GRASP) Ant Colony Optimization (ACO) Guided Local Search (GLS) Summary 1 Introduction Combinatorial optimization Examples • Knapsack, TSP, VRP, … A solution consisting of components Hard constraints Usually, in combinatorial optimization problems • e.g., VRP with pickup and delivery time windows General purpose metaheuristics with hard constraints Initial solution construction • Choose component one by one that gives feasible Tweaking • • • • To invent a closed Tweak operator To try repeatedly various Tweaks To allow infeasible solutions with distance from feasible one as quality To assign infeasible solutions a poor quality • Hamming cliff? 2 Introduction Components of solution e.g., edges between cities for TSP, pairs of jobs for T-problem Component-oriented methods Random selection of components • Greedy Randomized Adaptive Search Procedures (GRASP) • Algorithm 108 Favoring good components • Ant Colony Optimization (ACO) Punishing components related to local optima • Guided Local Search (GLS) 3 Ant Colony Optimization Two populations Set of components with pheromones as their fitness • e.g., all edges of TSP • Pheromone: historical quality of component Set of candidate solutions (ant trails) Free from Tweaking, possibly Algorithm 109 An Abstract Ant Colony Optimization Algorithm (ACO) 4 Ant Colony Optimization Ant System Algorithm 110 • The Ant System (AS) Selection of components based on desirability Initialization of pheromones • e.g., = 1, = popsize(1/C) where C is cost of tour constructed greedily Evaporation and update of pheromones Hill-climbing (optional) • Tweak, required Algorithm 111 • Pheromone Updating with a Learning Rate 5 Ant Colony Optimization Ant Colony System Changes from AS • • • • Elitist approach to updating pheromones Learning rate in pheromone updates Evaporating pheromones, slightly differently Strong tendency to select components used in the best trail discovered Algorithm 112 • The Ant Colony System (ACS) Elitist Component selection • With probability q, select component with highest desirability • Otherwise, do same as AS Disregarding linkage among components • Jacks-of-all-trade problem • c.f., N-population cooperative coevolution • Possible remedy: considering pairs of components? 6 Guided Local Search Avoiding some components for a solution Identifying components tending to cause local optima • Components that appear too often in local optima Penalizing solutions that use those components (toward exploration) c.f., Feature-based Tabu Search Fitness by quality and penalty (pheromone) Components whose pheromone is increased One with max. penalizability, in current solution Algorithm 113 Guided Local Search (GLS) with Random Updates • Detection of local optima? 7 Summary Combinatorial optimization Hard constraints Difficulties in construction of initial solution and Tweaking Component-oriented methods Randomly • e.g., GRASP Favoring with desirability • e.g., ACO Punishing with penalizability • e.g., GLS 8