A new ant colony optimization model for complex graph-based problems

Abstract

Nowadays, there is a huge number of problems that due to their complexity have employed heuristic-based algorithms to search for near-to-optimal (or even optimal) solutions. These problems are usually NP-complete, so classical algorithms are not the best candidates to address these problems because they need a large amount of computational resources, or they simply cannot find any solution when the problem grows. Some classical examples of these kind of problems are the Travelling Salesman Problem (TSP) or the N-Queens problem. It is also possible to find examples in real and industrial domains related to the optimization of complex problems, like planning, scheduling, Vehicle Routing Problems (VRP), WiFi network Design Problem (WiFiDP) or behavioural pattern identification, among others. Regarding to heuristic-based algorithms, two well-known paradigms are Swarm Intelligence and Evolutionary Computation. Both paradigms belongs to a subfield from Artificial Intelligence, named Computational Intelligence that also contains Fuzzy Systems, Artificial Neural Networks and Artificial Immune Systems areas. Swarm Intelligence (SI) algorithms are focused on the collective behaviour of selforganizing systems. These algorithms are characterized by the generation of collective intelligence from non-complex individual behaviour and the communication schemes amongst them. Some examples of SI algorithms are particle swarm optimization, ant colony optimization (ACO), bee colony optimization o bird flocking. Ant Colony Optimization (ACO) are based on the foraging behaviour of these insects. In these kind of algorithms, the ants take different decisions during their execution that allows them to build their own solution to the problem. Once any ant has finished its execution, the ant goes back through the followed path and it deposits, in the environment, pheromones that contains information about the built solution. These pheromones will influence the decision of future ants, so there is an indirect communication through the environment called stigmergy. When an ACO algorithm is applied to any of the optimization problems just described, the problem is usually modelled into a graph. Nevertheless, the classical graph-based representation is not the best one for the execution of ACO algorithms because it presents some important pitfalls. The first one is related to the polynomial, or even exponential, growth of the resulting graph. The second pitfall is related to those problems that needs from real variables because these problems cannot be modelled using the classical graph-based representation. On the other hand, Evolutionary Computation (EC) are a set of population-based algorithms based in the Darwinian evolutionary process. In this kind of algorithms there is one (or more) population composed by different individuals that represent a possible solution to the problem. For each iteration, the population evolves by the use of evolutionary procedures which means that better individuals (i.e. better solutions) are generated along the execution of the algorithm. Both kind of algorithms, EC and SI, have been traditionally applied in previous NP-hard problems. Different population-based strategies have been developed, compared and even combined to design hybrid algorithms. This thesis has been focused on the analysis of classical graph-based representations and its application in ACO algorithms into complex problems, and the development of a new ACO model that tries to take a step forward in this kind of algorithms. In this new model, the problem is represented using a reduced graph that affects to the ants behaviour, which becomes more complex. Also, this size reduction generates a fast growth in the number of pheromones created. For this reason, a new metaheuristic (called Oblivion Rate) has been designed to control the number of pheromones stored in the graph. In this thesis different metaheuristics have been designed for the proposed system and their performance have been compared. One of these metaheuristics is the Oblivion Rate, based on an exponential function that takes into account the number of pheromones created in the system. Other Oblivion Rate function is based on a bioinspired swarm algorithm that uses some concepts extracted from the evolutionary algorithms. This bio-inspired swarm algorithm is called Coral Reef Opmization (CRO) algorithm and it is based on the behaviour of the corals in a reef. Finally, to test and validate the proposed model, different domains have been used such as the N-Queens Problem, the Resource-Constraint Project Scheduling Problem, the Path Finding problem in Video Games, or the Behavioural Pattern Identification in users. In some of these domains, the performance of the proposed model has been compared against a classical Genetic Algorithm to provide a comparative study and perform an analytical comparison between both approaches.

Type
Antonio Gonzalez-Pardo
Antonio Gonzalez-Pardo
Associate Professor

Lecturer at the Computer Science Department. Main research interests are related to Computational Intelligence and Metaheuristics applied to Social Networks Analysis, and the optimization of graph-based problems.