Browse Theses

In this research, we propose two neural networks-based metaheuristic approaches for solving various well-known optimization problems in the literature. The first metaheuristic is called Augmented Neural Networks (AugNN). In this approach, traditional neural networks are augmented to allow embedding domain and problem-specific knowledge. The network architecture is problem specific and complex neural functions are used to (i) capture the constraints of the problem and (ii) apply a priority rule-based heuristic. Such an approach combines the advantages of known heuristics and the iterative learning of neural networks. We apply this approach to the classical open-shop scheduling problem (OSSP) and to the parallel schedule generation of resource-constrained project scheduling problem (RCPSP).

The second metaheuristic that we propose is the Adaptive Learning Approach (ALA). This approach employs a one-pass heuristic to give a good starting solution in the search space and uses a weight parameter to perturb the data of the original problem in order to obtain improved solutions. The weights are then adjusted employing a search strategy which involves reinforcement and backtracking. The random perturbation allows a non-deterministic local search. We apply the improvement-heuristic approach to the flow shop scheduling problem (FSSP) in conjunction with three well-known heuristics in the literature. We also use this approach for the serial schedule generation of the RCPSP and multi-mode resource constrained project scheduling problem (MRCPSP).

We empirically test our proposed approaches for each problem type on several benchmark instances in the literature and on some new problem instances generated in this study. We compare our results with lower-bound and best-known upper bound solutions. The results are extremely competitive with existing techniques such as genetic algorithms, simulated annealing, tabu search and ant colonies, in terms of both solution quality and computational times.