Identification of best sets of actions in Influence Nets

Article type: Research Article

Authors: Haider, Sajjad^{; *} | Zaidi, Abbas K. | Levis, Alexander H.

Affiliations: System Architectures Lab, George Mason University, Virginia 22030, USA

Correspondence: [*] Corresponding author. E-mail: [email protected]

Abstract: This paper presents a heuristic approach to solve the problem of best set of actions determination in Influence Nets. Influence Nets are a special instance of Bayesian Networks that model uncertain situations by connecting a set of desired effects to a set of actionable events through chains of probabilistic cause and effect relationships. Once an Influence Net is specified, a system analyst is often interested in identifying the set of action which has the highest probability of achieving a desired effect. The existing techniques to solve this problem, such as sensitivity analysis and exhaustive search, have limitations of their own. The proposed algorithm, named SAF, attempts to overcome these limitations. The paper also shows that the problem of best set of actions determination can be formulated as an instance of Mixed Integer Non Linear Programming (MINLP). An empirical study is presented that compares the performance of sensitivity analysis, SAF, and MINLP.

Keywords: Bayesian Networks, sensitivity analysis, Influence Nets, probabilistic reasoning, combinatorial optimization, mixed integer nonlinear programming

DOI: 10.3233/HIS-2008-5102

Journal: International Journal of Hybrid Intelligent Systems, vol. 5, no. 1, pp. 19-29, 2008