16 The Open Artificial Intelligence Journal, 2009, 3, 16-26 1874-0618/09 2009 Bentham Open Open Access Metaheuristic Algorithms for Task Assignment in Distributed Computing Systems: A Comparative and Integrative Approach Peng-Yeng Yin * , Benjamin B.M. Shao, Yung-Pin Cheng and Chung-Chao Yeh Department of Information Management, National Chi Nan University, 303 University Rd., Puli, Nantou 545, Taiwan Abstract: We consider the assignment of program tasks to processors in distributed computing systems such that system cost is minimized and resource constraints are satisfied. Several formulations for this task assignment problem (TAP) have been proposed in the literature. Most of these TAP formulations, however, are NP-complete and thus finding exact solutions is computationally intractable. Recently, some approximation methods like simulated annealing have been pro- posed, and simulation results exhibited the potential to solve the TAP using metaheuristics. In order to better understand the strengths and weaknesses of various metaheuristics applied to the TAP, we first propose two alternative metaheuris- tics—one using genetic algorithm and the other reinforcement learning algorithm—as well as their implementation de- tails. Extensive computational evidences of the two heuristic algorithms against that of simulated annealing are presented, compared and discussed. Based on these experimental results, a hybrid strategy employing both metaheuristics is then proposed in order to solve the TAP more effectively and efficiently. Keywords: Computer, heuristic algorithms, task assignment problem, distributed systems, genetic algorithms, reinforcement learning, simulated annealing. 1. INTRODUCTION A distributed computing system is defined as a collection of computers interconnected by a telecommunication net- work that attempts to disperse the data processing function and fits the needs of modern decentralized organization structures. Besides the capability of implementing a logically integrated information system for geographically dispersed corporations, distributed computing systems provide other benefits, such as quick access to data, higher system reliabil- ity, and ease of incremental growth [1]. On the other hand, efficient utilization of resources in distributed computing systems is also important [2]. In dis- tributed computing systems, it is characteristically difficult to assign the tasks of a program application to distributed processors such that a certain measure of system costs is minimized and system resources are effectively utilized. Several formulations of this task assignment problem (TAP) have been proposed in the literature to cope with various types of system costs and environmental constraints. In gen- eral, the TAP is NP-complete and finding exact solutions is computationally prohibitive [3]. In addressing this intractability issue, previous endeavors of TAP research can be classified into three areas. First, ex- act mathematical programming approaches using column generation [4] and branch-and-bound [5-6] have been pro- posed. Second, efficient algorithms have been developed for solving TAP on special computer architectures, such as lin- ear processor array, meshed processor graph, and partial k-tree *Address correspondence to this author at the Department of Information Management, National Chi Nan University, 303 University Rd., Puli, Nan- tou 545, Taiwan; Tel: +886-49-2910960; Fax: +886-49-2915205; E-mail: pyyin@ncnu.edu.tw communication graph [7-11]. Finally, metaheuristic algo- rithms like simulated annealing have been used to derive good enough approximate solutions within reasonable CPU time [12-13]. The current study belongs to the last domain of TAP re- search. Our primary objective is to provide a roadmap for better utilizing metaheuristic approaches and to incorporate multiple metaheuristics into one integrative framework for solving TAP more effectively and efficiently. Previous stud- ies have shown that the success of using metaheuristic algo- rithms depends on a proper administration of exploration and exploitation search in order to escape from local optimality [14]. Moreover, specific execution strategies may cause the same metaheuristic algorithm to behave differently. It is also noted that different metaheuristics have varying computa- tional performances for distinct applications and varied al- lowable execution time. Therefore, a hybrid approach incor- porating multiple metaheuristics may yield better perform- ance than a single approach [15]. It is advisable to conduct a thorough comparative study on these metaheuristics [16] so as to combine their favorable features into one comprehen- sive approach. In this paper, we first employ two metaheuristics-genetic algorithm and reinforcement learning-to compare with simulated annealing for solving the TAP. Then, by examin- ing their computational results, a hybrid algorithm combin- ing the two alternative metaheuristics is devised to solve the TAP more effectively and efficiently. The remainder of this paper is organized as follows. Section 2 formulates the TAP to be addressed in this paper. In Section 3, two metaheuris- tics based on genetic algorithm and reinforcement learning are proposed individually. Section 4 then follows and pre- sents the comparative simulation results, based on which a