An investigation of Monte Carlo tree search and local search for course timetabling problems

The work presented in this thesis focuses on solving course timetabling problems, a variant of education timetabling. Automated timetabling is a popular topic among researchers and practitioners because manual timetable construction is impractical, if not impossible, as it is known to be NP-hard. A two-stage approach is investigated. The first stage involves finding feasible solutions. Monte Carlo Tree Search (MCTS) is utilized in this stage. As far as we are aware, it is used for the first time in addressing the timetabling problem. It is a relatively new search method and has achieved breakthrough in the domain of games particularly Go. Several enhancements are attempted on MCTS such as heuristic based simulations and pruning. We also compare the effectiveness of MCTS with Graph Coloring Heuristic (GCH) and Tabu Search (TS) based methods. Initial findings show that a TS based method is more promising, so we focus on improving TS. We propose an algorithm called Tabu Search with Sampling and Perturbation (TSSP). Among the enhancements that we introduced are event sampling, a novel cost function and perturbation. Furthermore, we hybridize TSSP with Iterated Local Search (ILS). The second stage focuses on improving the quality of feasible solutions. We propose a variant of Simulated Annealing called Simulated Annealing with Reheating (SAR). SAR has three features: a novel neighborhood examination scheme, a new way of estimating local optima and a reheating scheme. The rigorous setting of initial and end temperature in conventional SA is bypassed in SAR. Precisely, reheating and cooling were applied at the right time and level, thus saving time allowing the search to be performed efficiently. One drawback of SAR is having to preset the composition of neighborhood structures for the datasets. We present an enhanced variant of the SAR algorithm called Simulated Annealing with Improved Reheating and Learning (SAIRL). We propose a reinforcement learning based method to obtain a suitable neighborhood structure composition for the search to operate effectively. We also propose to incorporate the average cost changes into the reheated temperature function. SAIRL eliminates the need for tuning parameters in conventional SA as well as neighborhood structures composition in SAR. Experiments were tested on four publicly available datasets namely Socha, International Timetabling Competition 2002 (ITC02), International Timetabling Competition 2007 (ITC07) and Hard. Our results are better or competitive when compared with other state of the art methods where new best results are obtained for many instances