Abstract
This study proposes Neural Network frameworks for solving two different problems, namely--Scheduling and Bankruptcy Prediction.
In the scheduling problem, tasks with precedence relationship are to be assigned to machines, which may be identical or non-identical, such that the makespan time is minimized. There are arbitrary number of machines and task preemption is not allowed. In the proposed framework, a scheduling problem is represented in the form of a network resembling a Neural Network. It is an iterative procedure wherein each iteration produces a feasible solution, and a learning rule improves the solution from one iteration to the next, much like in a Neural Network. The proposed framework is implemented in Pascal running on microcomputer and tested on 200 problems for the case of identical machines and 124 problems for the case of non-identical machines. For the case of identical machine, the results are compared with those of a known heuristic and we find that in 77% of the cases, the proposed framework gave improved solutions. The improvement in makespan was 1.13% on average. For the case of non identical machines, the proposed framework improved over heuristic in 60% of the cases and the improvement in makespan was 3.69% on average.
For the Bankruptcy Prediction Problem, we propose a framework which allows us to detect trends across time in the financial data of companies and accordingly discriminate between healthy and failing companies. The proposed Neural Network framework is implemented in Pascal running on microcomputer and tested on companies from four different SIC Code groups. Results were compared with those obtained by using existing statistical techniques, namely--multivariate discriminant analysis (MDA) and logistic regression (Logit) and it was found that the proposed framework performed better than Logit, which is the better of the two statistical techniques, at a p-value of 0.062. The Neural Network framework was also used to detect bankruptcy using single year data. The results were compared with those obtained by Logit and MDA and it was found that the Neural Networks outperformed logit for all the SIC Codes at significant p-values.