Deep Reinforcement Learning Method for Building Layout Generation under Complex Terrain Conditions

Under the complicated condition of terrain, there are challenges for building layout generation, including the irregularities of the terrain and the multi-objective constraints. This study is proposing a solution using Multi-Agent Deep Reinforcement Learning (MADRL). By modeling each building instance as an individual agent, the distributed decision-making framework is formed through deep Q network (DQN), in which each agent dynamically adjust the center coordinates of each building according to the discrete operation space (1 meter up, down, left, right, no movement) and at the same time, the data of the terrain elevation, the building boundary constraints and the spatial relationship of the adjacent buildings are combined through the state representation and formed a multi-dimensional state vector with position coordinates, constraints satisfaction of the spacing and the adaptability of the terrain. In order to solve the problem of late constraint failure caused by the early layout errors, a curriculum learning strategy is introduced. The reward function design uses a multi-objective weighted mechanism and compromises between the conflicting objectives and adopts a dynamic weight adjustment mechanism. The MADRL method has the maximum building coverage of 76.5%, the average distribution uniformity index of 0.761, and improved layout connectivity index. This paper presents an efficient, flexible, and adaptable design solution for the construction of buildings in the complex terrain condition.