Abstract
In architectural planning stage, multi-agent system based on human behavior models is considered to be effective in order to estimate space characteristics of objective buildings such as efficiency, amenity, safety, and so on. In multi-agent simulations, at first, human behavior rules are set up, and the efficiency of the objective space is estimated by observed results on agents' behaviors with employed rules. However, it is difficult to configure human behavior rules, because actual human behaviors are complex. Therefore, acquisition methods of human behavior rules in accordance with objective spaces and/or issues are considered to be an important problem.
In this paper, aiming to develop adaptable agent models, artificial neural network (ANN) is employed as acquisition and control methods of agents' behaviors. As for training methods of ANN, genetic algorithm (GA) is employed in order to determine optimal weights and biases of ANN. As for objective spaces, passage spaces with complicated shapes are employed. Based on simulation results, the effectiveness of acquired behavior rules by GA, the generalization capabilities of ANN, and the effectiveness of multi-agent simulations are discussed. Furthermore, possibilities on acquisition methods of adaptable human behavior models are also discussed.
