Energy and Buildings 127 (2016) 714–729 Contents lists available at ScienceDirect Energy and Buildings j ourna l ho me pa g e: www.elsevier.com/locate/enbuild Improving evolutionary algorithm performance for integer type multi-objective building system design optimization Weili Xu ∗ , Adrian Chong, Omer T. Karaguzel, Khee Poh Lam Center for Building Performance and Diagnostics, Carnegie Mellon University, Pittsburgh, PA 15213, USA a r t i c l e i n f o Article history: Received 16 March 2016 Received in revised form 14 June 2016 Accepted 14 June 2016 Available online 21 June 2016 Keywords: Multi-objective evolutionary optimization Building system design Building cost estimation Optimization performance a b s t r a c t Building system design optimization is becoming popular for design decision making. State-of-the-art technique that couples evolutionary algorithms with a building simulation engine, which is time con- suming and often cannot reach the “true” optimal solutions. Studies addressing these issues focus on implementing strategies such as fine tuning optimization algorithm’s parameters, hybrid evolutionary algorithms with a local search algorithm or optimizing meta-models. Unlike the previous studies, this paper proposes two improvement strategies for building system design optimization. The two strategies, adaptive operators approach and adaptive meta-model approach, modify the behaviors of conventional evolutionary algorithms to improve the optimization convergency and speed performance. To demon- strate the effectiveness of these two strategies compared to conventional algorithms, a case study was conducted. The case study observed high convergency performance from both strategies with 30% and 60% time savings respectively. Furthermore, this study examines the performance comparison in respect to convergency, diversity preservation and speed between these two strategies. © 2016 Elsevier B.V. All rights reserved. 1. Introduction In building design process, decisions are usually constrained by multiple factors. Solutions that satisfy building owners’ require- ments are commonly infeasible to reach by performing parametric studies. Therefore, more efficient methods are needed in search- ing the solution space to find optimal solutions, which not only reduce the evaluation time, but also provide optimal designs to achieve building owners’ investment goals. This raises the topic of building system design optimization (BSDO) that uses advanced optimization algorithms for searching for optimal design solutions. The topic has increasingly drawn attention from the academic com- munity, as well as the architecture, engineering and construction (AEC) industry. Current studies well address the behavior of var- ious optimization algorithms including pattern search methods and stochastic methods [1]. These studies have also extensively examined various objectives and levels of detail regarding single system performance optimization and integrated building design optimization [2]. In recent years, both academia and industry are developing tools that support the ease of implementing optimiza- tion in building design process. Tools such as MOBO [3], GenOpt [4] and jEPlus [5] provide user-friendly interface and capability of ∗ Corresponding author. E-mail address: weilix@andrew.cmu.edu (W. Xu). coupling building design evaluation toolset (TRNSYS, EnergyPlus, etc.) as well as allow stakeholders to explore their design options more effectively. Although BSDO has been actively discussed among academic community for decades, it is still not a common technique used in today’s typical building projects. One of the barriers is computation time. A typical BSDO process could take days to find optimal solu- tions. In order to reduce the computation time, many researchers are looking for computational efficient strategies which can fully utilize computational resources to boost the optimization speed [6]. Although these researches did not address the computation- ally expensive design evaluation process in BSDO, the strategies proposed can effectively alleviate the impact of evaluation speed. Such strategies can be categorized into three types, namely: par- allel computing, model simplification and meta-model approaches [1]. Parallel computing allows optimization algorithms distribut- ing a number of simulation tasks into multiple process threads simultaneously, thus reducing the overall computation time. Implementation of this approach requires advance level of pro- gramming skills, nevertheless, the majority of current energy simulation software have included such features. Simplifying the complexity of problems is another popular approach that fre- quently appears in many studies. Such studies usually construct a simple geometry layout with small amount of design variables. However, simplification highly relies on expert knowledge and designers have to take risks for losing building system interaction http://dx.doi.org/10.1016/j.enbuild.2016.06.043 0378-7788/© 2016 Elsevier B.V. All rights reserved.