This paper presents a new algorithm for enhancing the efficiency of simulation-based optimisation using local search and neural network metamodels. The local search strategy is based on steepest ascent Hill Climbing. In contrast to many other approaches that use a metamodel for simulation optimisation, this algorithm alternates between the metamodel and its underlying simulation model, rather than using them sequentially. On-line learning of the metamodel is applied to improve its accuracy in the current region of the search space. The proposed algorithm is applied to a theoretical benchmark problem as well as a real-world manufacturing optimisation problem and initial results show good performance when compared to a standard Hill Climbing strategy.