Architectural planning with shape grammars and reinforcement learning: Habitability and energy efficiency

Abstract

This paper describes the generation of sketches of small single-family dwellings that satisfy habitability requirements and are energy efficient. The proposed approach considers three stages in the generation process, and each one is based on a combination of shape grammars and reinforcement learning. First a set of very simple shape grammar rules is defined that are capable of generating a great variety of sketches. In order to guarantee the generation of sketches that are both suitable for habitation and energy efficient, a reinforcement learning process is applied on this set. Then the grammar so trained is used to generate only “good” sketches. More precisely, the learning process applies positive rewards to sketches that satisfy desired habitability and energy efficiency guidelines. As a result, sequences of grammar rules that lead to good sketches are identified.In this paper we present the general approach followed to develop the system and describe in detail the procedure applied in the reinforcement learning process. Experimental results are also presented, to show convergence of the learning process, and to compare the obtained results with those of real designs. A standard energy simulation program is used to validate the approach.