Agent Based Evolutionary Dynamic Optimization

Abstract

Summary Agent-based Evolutionary Search (AES) has attracted a growing interest from the evolutionary computation community in recent years due to its robust ability in solving large scale problems, ranging from online trading, disaster response, modeling social structures to financial investment planning. In order to solve these problems, a great variety of intelligent techniques have been developed to improve the framework and efficiency of AES. This chapter investigates an agent based evolutionary search algorithm in which the agents are updated and co-evolve to track dynamic optimum by imitating the exhibited feature of living organism. In the proposed algorithm, all agents live in a lattice like environment, where each agent is fixed on a lattice point. In order to increase the predefined energy function, individual agent can compete with its neighbors and also can acquire knowledge through cumulative information. For the purpose of maintaining the diversity of the population, random immigrants and adaptive primal dual mapping schemes are incorporated. Simulation experiments on a set of dynamic benchmark problems show the proposed AES algorithm can yield a better performance on dynamic optimization problems (DOPs) in comparison with several peer algorithms.

Multiagent evolutionary computation for complex problems

“…As a new paradigm, MAEC becomes a hot topic to solve a range of complex problems, such as global numeric optimization [72,78,[84][85][86][87], reactive power dispatch [73], rule-based knowledge extraction [74], constraint satisfaction [75,79], combinatorial optimization [76], dynamic optimization [77], image feature extraction [80,81], 7000-queen problems [82], server job scheduling [88], robot soccer in Keepaway [89,90], iterated prisoner's dilemma [91,92], trust management in wireless senor networks (WSNs) [95], mobile ad hoc networks (MANETs) [96] and vehicular ad hoc networks (VANETs) [97], fashion design problems [100], traffic congestion forecasting [32][33][34]70], etc.…”

“…In the first category "agent based EC", most approaches only utilize part of characteristics of intelligent agents (e.g., local views of agents) [72][73][74][75][76][77][78][79][80][81][82][83]. In addition, many methods design special actions for the particular problems, which limit their generality on other problems [78][79][80][81][82][83].…”

“…At the same time, all agents have local views and live in, for example, the lattice-like environment [72][73][74][75][76], two dimensional grid environment [77], ring-like environment [146], etc. In these environments, every agent is fixed on a relatively stable position and only has local views to adjacent regions.…”

“…In [72][73][74][75][76][77], there are four types of predefined actions for agents: the neighborhood competition operator, the mutation operator, the advanced evolutionary operator (e.g., orthogonal crossover operator [126]) and the self-learning operator (e.g., local search [72,73], learning table [76], statistics based learning behavior [77]). Basically, these actions are traditionally evolutionary operators with specified parameter settings, which are proposed and verified as effective operators to solve particular problems.…”

“…For instance, the action of "orthogonal crossover operator" is controlled by the predefined crossover probability, which is set as a fix value (e.g., p c = 0.1 [72]) before the algorithm starts. Thus, these agents can be classified into the simple reflex agents [72][73][74][75][76] or the model-based reflex agents [77].…”