An agent-based framework for problem solving in symbiotic simulation systems

• Main Author: Aydt, Heiko
• Other Authors: Cai Wentong
• Format: Theses and Dissertations
• Language: English
• Published: 2011
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Simulation and modeling; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/46291
• Institution: Nanyang Technological University

Symbiotic simulation can be used for problem solving processes in the context of applications concerned with real-world physical systems. These kind of systems are often highly complex and exhibit non-linear behaviour which necessitates the use of simulation techniques to evaluate possible solutions to problems concerned with these systems. Problem solving can be described as an optimisation process concerned with the minimisation/maximisation of one or more objectives (expressed in terms of performance indicators). Optimisation requires the use of knowledge regarding the problem in order to effectively and efficiently solve it. Although black-box optimisation (i.e., without use of any knowledge regarding the problem) can be performed, it is easily outperformed by algorithms that incorporate domain knowledge. The more knowledge about a certain problem is incorporated into an algorithm, the more specialised the algorithm becomes. While specialisation generally improves the performance of an algorithm in solving a particular problem, it does so only at cost of decreasing re-usability of the algorithm for other problems. This is due to the implications of the no-free-lunch theorems. An autonomous problem solver agent which is meant to replace a human problem solver and automatically perform what-if analyses, needs to be able to solve different problems during its life span. This requires a flexible approach that does not statically hard-code information about a problem into the problem solving algorithm. Instead, it is necessary to dynamically incorporate problem-specific knowledge. In this dissertation we address this issue and establish a framework for constructing problem solver agents, based on symbiotic simulation. The approach in this dissertation is three-fold. First, we establish a theory on symbiotic simulation which also takes consideration of related work. As a result we propose a taxonomy on symbiotic simulations and introduce various classes of symbiotic simulation systems. Based on this taxonomy we can clearly define problem solving in symbiotic simulation and specify the what-if analysis process. Second, we argue for the use of evolutionary computing and propose our method for separating problem specific knowledge from the implementation of an evolutionary algorithm using an appropriate language.