Software Engineering methodologies have demonstrated their importance in the efficient solution of complex real-world problems. The process of software development can be viewed as searching through the state space of all possible programs. Evolutionary computation methods are useful in this search process due to their higher level of complexity. We are interested in performing an efficient search through the leverage is of Software Engineering techniques in order to maintain detailed information about program constraints. Our goal is to focus the search through identification of these constraints.; This thesis takes software testing methodologies and applies them to software design. Software testing processes reinforce and verify the design by the practice of determining program faults through the identification of knowledge that can allow the programmer to pin-point its cause and relate them back to the specification. We rely on complementary approaches in Software testing which are white box and black box testing. White box testing examines a programs structure while black box examines outputs in relevance to input data sets. These are applied in the context of software design in which the white box is first applied in order to generate a prototype. Once the program has been developed to a suitable level of performance, a black box approach is applied. This process runs in sequence until a suitable solution is found.; We apply these testing concepts through the utilization of Cultural Algorithms. Cultural Algorithms enhance the evolutionary process through the application of a belief structure to the traditional evolutionary approach. Our approach two Cultural Algorithms with one focusing on white box and the second on black box. This is termed as a Dual Cultural Algorithms with Genetic Programming. We apply this to a benchmark problem, the quadratic equation, which has initially been used by Zannoni and Reynolds [Zannoni 1996]. Here, we present a more effective approach in program generation in comparison to a standard GP approach. The solutions generated are also demonstrated to less complex than those generated with standard GP approaches.; Next, we apply this to a multi-agent system developed in order to simulate transactions in a durable goods market. Here, we find that a near-optimal strategy has a diminishing effect when heterogeneous factors are applied to our agents. We utilize the DCAGP framework to calibrate our agent-based model by allowing it to utilize the multi-agent system by allowing the evolutionary framework to use the multi-agent system as a performance function. This approach allows us to produce a near optimal solution in less generations than standard genetic programming methodologies.
展开▼
