Search with evolutionary ruin and stochastic rebuild: A theoretic framework and a case study on exam timetabling

•This paper presents a new search method which tries to learn and adapt to the changing environments during the search.•It improves the original R&R idea by implementing a phase of Evolutionary Ruin to mimic an evolution within single solutions.•It achieves optimisation by an iterative process of component evaluation, solution disruption and stochastic constructive repair.•It presents a formal framework and implements a Markov chain analysis on some theoretical properties of the approach.•Experimental results on exam timetabling benchmark problems have demonstrated the strength of the approach.

This paper presents a state transition based formal framework for a new search method, called Evolutionary Ruin and Stochastic Recreate, which tries to learn and adapt to the changing environments during the search process. It improves the performance of the original Ruin and Recreate principle by embedding an additional phase of Evolutionary Ruin to mimic the survival-of-the-fittest mechanism within single solutions. This method executes a cycle of Solution Decomposition, Evolutionary Ruin, Stochastic Recreate and Solution Acceptance until a certain stopping condition is met. The Solution Decomposition phase first uses some problem-specific knowledge to decompose a complete solution into its components and assigns a score to each component. The Evolutionary Ruin phase then employs two evolutionary operators (namely Selection and Mutation) to destroy a certain fraction of the solution, and the next Stochastic Recreate phase repairs the “broken” solution. Last, the Solution Acceptance phase selects a specific strategy to determine the probability of accepting the newly generated solution. Hence, optimisation is achieved by an iterative process of component evaluation, solution disruption and stochastic constructive repair. From the state transitions point of view, this paper presents a probabilistic model and implements a Markov chain analysis on some theoretical properties of the approach. Unlike the theoretical work on genetic algorithm and simulated annealing which are based on state transitions within the space of complete assignments, our model is based on state transitions within the space of partial assignments. The exam timetabling problems are used to test the performance in solving real-world hard problems.