Automation in Construction 14 (2005) 33 – 44 www.elsevier.com/locate/autcon
Decision support and design evolution: integrating genetic algorithms, CFD and visualization Ali M. Malkawia,*, Ravi S. Srinivasana, Yun K. Yia, Ruchi Choudharyb a
Department of Architecture, School of Design, University of Pennsylvania, 207 Meyerson Hall, 210 South 34th St., Philadelphia, PA 19104-6311, USA b College of Architecture, Georgia Institute of Technology, Atlanta, USA Accepted 19 June 2004
Abstract This paper describes a decision support evolution model using Genetic Algorithm (GA) as the evolution algorithm and Computational Fluid Dynamics (CFD) as the evaluation mechanism. The model is integrated with a visualization module to allow users to interact and select form instances as the design evolves. The advantages of such an evolutionary decision support design approach is that diverse instances of the state space can be investigated in relation to specific goal requirements which will enhance the possibility of discovering a variety of potential solutions. The model allows the user to explore and visualize the design evolution and its form generation in an attempt to stimulate the designer creativity that might contribute to their output. The process uses an iterative approach that allows design to be evaluated using CFD analysis automatically to maximize several thermal and ventilation criteria. Design change will then be performed, remeshed and displayed based on the evolutionary algorithms. The process allows the user to experience the morphing of the design based on its performance and continues until the designer has the opportunity to visualize the evolution of the final set of design alternatives. D 2004 Elsevier B.V. All rights reserved. Keywords: Decision support; Design evolution; Genetic algorithms; CFD; Visualization
1. Introduction and background
* Corresponding author. Tel.: +1 215 898 5728; fax: +1 215 573 2192. E-mail addresses:
[email protected] (A.M. Malkawi)8
[email protected] (R.S. Srinivasan)8
[email protected] (Y.K. Yi)8
[email protected] (R. Choudhary). 0926-5805/$ - see front matter D 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.autcon.2004.06.004
Thermal and ventilation performance criteria as a mean to generate creative design alternatives require an in-depth knowledge of the problem and considerable expertise. The complexity of this is due in part to the difficulty of evaluating the performance of such problems that are affected by many variables,
34
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44
and translate these evolutions into meaningful design ideas. For example, although within conditioned spaces, thermal and ventilation evaluation are significantly affected by air distribution, which is controlled by the supply flow rate, yet it is difficult to predict the resulting airflow distribution. This is due to the fact that the resulting distribution of air is dependent on many factors such as the type, size and location of the air distribution device, space geometry, heat transfer across space boundaries, and internal objects such as furniture, equipment and the occupants, etc. Through the introduction of CFD as a simulation mechanism, the opportunity to evaluate flows in spaces at high precision levels—something that is too cumbersome to do through experimentation and not possible by lower level simulations, can be achieved. Several projects have demonstrated the power of CFD for evaluating the building thermal environment during the past decade such as natural ventilation design [1,2]; building energy and thermal comfort simulations [3–5]; particulate dispersion in indoor environment [6], indoor air quality assessment [7–9]; prediction of fire and smoke in buildings [10]. To conduct the evaluation and possible optimization of the building thermal environment, parametric analyses are used. Parametric studies involve solving the problem several times with different sets of parameter variables such that solutions can be found. The effect of the parameter variables upon the design problem can be studied and evaluated. In addition, by varying key design parameters, and comparing the resulting design solutions, tradeoff between the different values of the parameter variables may be sought. Many projects have demonstrated the usefulness of parametric studies with CFD simulations both for identifying the key parameters in the simulation as well as for exporting different design alternatives [11–15]. One of the main limitations associated with parametric studies that involve CFD simulations is that they are time consuming and tedious. It is extremely cumbersome to build a CFD simulation for every change in the value of a parameter: each design change made to the space geometry, supply terminals or space boundaries requires the CFD user to remodel, remesh and subsequently, recompute the airflow. Additionally, the complexity of relationships
between a set of design parameters and the resulting ventilation performance is often hard to track. Tradeoff between different design changes generally remains obscured. Thus, the design space is difficult to explore systematically and solutions can be overlooked. Evolutionary techniques coupled with visualization can offer one solution to such a problem that will provide automation of optimization and interactive user intervention. Evolutionary algorithms, one of the byproducts of artificial intelligence field, have traditionally been used to solve optimization problems. In addition, they can be used for design decision making [16– 18]. The evolutionary approach is a bgenerate-andtestQ approach which corresponds well to the procedures for design synthesis and evaluation in the design process. In an evolutionary-based generative process, design representations are specified as a set of parameters, and a corresponding set of constraints to achieve optimization. Optimization is a formal design paradigm defining a rigorous and efficient process for iteratively searching best possible decisions, given a set of goals and circumstances in which those goals must be achieved [19]. Within its framework, design is cast in a goal oriented decision-making model. Goals are defined by desired building performance, and a causal relationship is presumed between design decisions and goals. Due to this causality presumption, which implies that design decisions are essentially driven by performance feedback, design optimization is a natural formalization of the simulation-based design process in buildings. It improves design performance by systematically searching for design decisions that meet stated goals, and thereby facilitates efficient and purposeful use of simulation tools in design decision making. With the advent of new technologies in the field of design evolution and optimization, the designer’s role shifts from that of a creator of individual instances of a style to that of a meta-designer or creator of an entire style of family [20]. Such generative designs assist in the rapid generation of computable design descriptions, design models or design representations [21]. In addition, these design evolutions can be used as an aid in stimulating the designer creativity [22]. The advantage of such an evolutionary approach is the creation of diverse
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44
sections of the state space that increases the possibility of discovering a variety of potential solutions [23]. Recently, several research projects were conducted that took advantage of design optimization [24–28]. Although these studies illustrated variety of approaches that target optimizing design variables for variety of evaluation criteria, they did not attempt to capture all possible design variable evolutions which provide larger search space for the designers to interact with. This paper presents a decision support design evolution model that allows efficient exploration of design alternatives. The goal is to automate portion of the analysis process and utilize techniques that will aid designers in making decisions. The system developed allows efficient exploration and visualization of design alternatives that maximize performance ventilation requirement. The design evolution uses a GA module that generates the shape of the design instances. The process proceeds in an iterative manner. For each design change, a CFD analysis is automatically performed to evaluate the thermal and ventilation efficiency, and the results are fed back into the GA module. This process runs recursively till the best designs are returned. To capture the system’s solutions and in-between design instances as the solution is evolving, the system is integrated with a morphing module. This module allows the diverse instances of designs to be visualized along with their performance in order to enhance the user’s potential of design discovery and provide aid within the designanalysis process.
2. The decision support design evolution structure The decision support design evolution model consists of a four-layer structure (Fig. 1). It uses GA as the evolution algorithm and CFD as the evaluation mechanism. The model runs recursively till the best designs are identified. This eliminates the need for the user to specify, design and arrange a display each time an output is required. All possible design evolutions can be visualized through a morphing module and their performance can be predicted. While the GA generates new forms based on predefined decision goals, the morph module allows
35
Fig. 1. Four-layer decision support design evolution model.
the visualization of all possible space geometries between two generated forms. During morphing of two consecutive forms, users can intervene and select an instance of the changing room geometry, while the GA continues its search. The selected design instance can then be evaluated for its performance using an integrated CFD engine. Temperature and velocity contours are automatically generated for users to visualize and checked for performance criteria. This evolutionary design approach creates discrete instances of designs that satisfy the performance targets by providing larger search space for designers to interact with through the integration of GA, CFD and morph visualization (Fig. 2). To implement the model, a formal representation of the problem was specified. This included a statement of goals (performance requirements), design decisions for achieving those goals, and a mathematical description of any requirements that the design must satisfy. Design decisions are examined against goals using an analysis (CFD simulation) model. The analysis model takes values of design decisions and parameters as input, and returns corresponding values of performance goals as
36
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44
Fig. 2. Decision support evolution model.
output where GA will drive the search for new forms and the morphing module will be used to visualize how the search is evolving.
The model is implemented to allow three-dimensional orthogonal geometries to be presented (Fig. 3). The space is defined by length, width and height and
Fig. 3. Representation of room.
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44 Table 1 Performance and geometry targets represented as variables and parameters Performance targets Target temperature Supply temperature Supply velocity Geometry targets Room length Room width Room Height Window location
Variables
Parameter(s)
S Tmax, S Tmin S Vmax, S Vmin
TT ST SV
R Lmax, R Lmin R Wmax, R Wmin R Hmax, R Hmin
Window left distance Window right distance Window floor distance Window ceiling distance Supply duct location
W Lmax, W Lmin W Rmax, W Rmin W Fmax, W Fmin W Cmax, W Cmin
Supply X distance Supply Y distance Supply length Supply height Return duct location
S Xmax, S Xmin S Ymax, S Ymin S Lmax, S Lmin S Hmax, S Hmin
Return Return Return Return
R Xmax, R Xmin R Ymax, R Ymin R Emax, R Emin R Imax, R Imin
X distance Y distance length height
RL RW RH W Wx , where x*=n, e, w, s, f, r WL WR WF WC S Wx , where x*=n, e, w, s, f, r SX SY SL SH R Wx , where x*=n, e, w, s, f, r RX RY RE RI
contains a window, a door, one supply duct and an extract terminal. The window is located on the south wall surface of the space. It is defined by two width and two height variables for the distance of each window edge from the nearest wall edge. The window size is computed by subtracting the sum of the two variables from the overall width and height of the south wall. The door to the space is a fixed parameter, and is centrally located on the north wall surface of the space. The space is allowed to have two supply and two extract terminals. The terminals are represented by three sets of design variables: the wall on which a duct is located, its position on the wall, and the effective area of the grille. Supply velocity is computed on the basis of space volume, area of the supply ducts and air change per hour required for the space. The problem constitutes 21 design variables: 4 discrete and 17 continuous. Each variable is asso-
37
ciated with explicit upper and lower limits and the interval over which it can be incremented. The design representations are specified as a set of variables and parameters (Table 1). The variable and parameter sets form the genetic elements of the design and define the shape of the design instances. Variables are design attributes that affect design behavior and are varied to find values that best satisfy the evaluation criteria. The term parameter is used to describe other design attributes that also affect design behavior but are assigned fixed values that are not varied during the optimization. The space is evaluated for minimizing temperature and velocity differences from a specified temperature and velocity magnitude in the occupied zone. Numerically, the occupied zone is the region 2 ft away from all walls and 3 ft from the ceiling (Fig. 4). 2.1. Design evolution The selection of a suitable evolutionary algorithm depends on the problem formulation and what one requires from the optimization. For example, gradient-based methods will be fast and rigorous for solving continuous problems, and a large combinatorial problem can be best handled using stochastic methods. Over the past few years, stochastic methods such as Simulated Annealing and Genetic Algorithms have also become very popular, and have been applied to a range of problems for optimizing thermal and lighting performance based on building enclosure, HVAC design, and control schedules [29–32].
Fig. 4. The occupied zone of the room.
38
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44
They are attractive mainly because they solve a wide range of problems and do not require functions to be smooth and can handle mixed-discrete variables. Since building design analysis is a mixed variable problem, GAs are used to explore the design space. Mixed variable problems includes continuous variables such as window sizes, and discrete such as wall choices, etc. GA is a one of the stochastic methods of optimization, which involve random elements in their search strategies. The genetic algorithm used was the steady-state GA (GALib, a C++ genetic algorithm library). Steady-state GA use overlapping populations, which means that it allows the best of a population to be carried over to the next generation [33]. The use of GAs requires the decision variables to be encoded as a chromosome. The algorithm randomly generates an initial population of chromosomes or begins with a user-supplied population, and then applies genetic operators to the best chromosomes in that population to generate dchildrenT or new genotypes. A new population is created by choosing the best chromosomes from both the old and newly generated chromosomes by applying selection criteria. This process is repeated for specified number of iterations/generations, after which the algorithm terminates. The best chromosome(s) from the final generation is output as the optimum solution—there are many variants to this idea [34], including probabilistic selection of the bbestQ chromosome(s), selection from only the child chromosomes, etc. The GA can be designed to search in many different ways. The success of the algorithm heavily depends on how the genetic operators are defined, which individuals (and how many) are selected for genetic operations, what selection criteria are applied for forming subsequent populations, and how other parameters are set. There are no hard rules that dictate these decisions. For this work, the design variables are represented as an array of allele sets or dspace genomeT (Fig. 5). Allele sets of a variable contain values that the variable can assume. This means that each element of the space genome is mapped by using a value in its corresponding allele set. For example, if the first allele set is {10, 25}, the first element of the genome will be within this set. Depending on the variable, allele sets have been added in different ways, (enumerated, bounded or bounded with discretiza-
Fig. 5. Genome allele set representing the room.
tion), to the elements of the space genome. The advantage of this approach is that since the variable bounds are explicitly defined, the number of infeasible designs that would be otherwise generated is substantially reduced. Apart from explicit upper and lower bounds, other constraints have been included to ensure the feasibility of the design before it is evaluated. These include constraints on the window size and constraints on the location of supply and extract terminals to prevent overlapping. Since GALib does not handle constraints explicitly, the window size and overlapping constraints are modeled as penalty functions, and included in the design evaluation stage as the first design evaluation step. Two operators were utilized for mutation, flip mutation and one point crossover. The set of rules defining mutations is called genetic operators. A crossover is analogous to mating in natural evolution, and
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44
corresponds to swapping decision variables between two existing chromosomes to create two new chromosomes. Mutation changes some decision variables in an existing chromosome, thereby generating a new chromosome. The GA generates a space genome and the design is passed for evaluation. During this evaluation process, design is first evaluated for constraint satisfaction. This is an important step because even if one design constraint is violated, the design cannot be modeled and meshed correctly for the CFD analysis. Therefore, when a constraint is violated, it is penalized heavily and the penalty value is returned as the fitness value (objective function value) of the design. In this way, it is ensured that the design will be rejected in the next generation. Once a feasible design is found, a three-dimensional model is constructed to be evaluated by CFD. The results obtained from the evolutionary algorithm were evaluated by cross-comparing them with results obtained from solving the problem by a gradient-based algorithm. A side-wall supply problem with three continuous variables [35] was modeled as a test case. This test problem constitutes a single sidewall supply terminal for an air-conditioned room of fixed area, thermal load, and boundary conditions. The design variables included position of the supply on the smaller width of the wall, the length of the supply duct and supply velocity. Constraints are posed on the throw length of the air jet and on the distance of the supply from the ceiling. The objective function was formulated for minimizing deviation from the target mean velocity in the occupied zone. The test case was modeled and solved using genetic algorithms and a gradient-based optimizer. The problem converged to the same solution in both cases: 8% of the total nodes deviated from the target mean velocity range in the occupied zone.
39
Fig. 6. Layer 2—performance evaluation.
providing a penalty value that is returned back to layer 1. This value is used to reject the evolution of such future instances and to ensure the generation of different designs. If CFD analysis converges to a solution, temperature and velocity measurements— computed as a function of supply, extract and local air characteristics—at each node within the occupied zone are read to a file. The value at each node is evaluated for its deviation from a set range of temperature and velocity. The GA evaluates each design by its fitness value. Therefore, each time the value at the node deviates by more than a certain percentage, the fitness-function value decrements. The maximization of the fitness function value therefore becomes equivalent to minimizing the total deviation value. After evaluating all the nodes for temperature and velocity magnitude, the final fitness value is returned to the GA and the form is sent to layer 3 for further morphing. The objective function, in this case, is to minimize deviation from desired velocity in the occupied zone. Objective function: Minimize f ð xÞ ¼ jNt Ni j
2.2. Performance Evaluation The Performance Evaluation layer uses CFD to perform the thermal analysis. Gambit v2.0.4 is used for mesh generation followed by Fluent v6.0 (Fig. 6). If there is a physical configuration error or inappropriate simulation setup, the CFD does not converge, prompting the design to be penalized. This is accomplished, as discussed earlier, by
where i is the room node being evaluated N t—target velocity in the occupied zone. The process is repeated for bnQ number of generations for a specified population size and GA returns the best design(s) from the current population. The best design(s) can be evaluated interactively, through layer 3, for design tradeoff and designer preferences.
40
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44
Fig. 7. Layer 3—morph visualization.
2.3. Morph Visualization The Morph Visualization layer aids in the design evolution process by visualizing the intermediate designs, which may, in normal circumstances be difficult to locate (Fig. 7). Morphing is an integral part of design evolution since it forms a smooth transition between pairs of consecutive key frame models. The morphing module generates design instances between two consecutive forms analyzed by the CFD (between F i+1 and F i ) to evolve new forms ( F mi ) through bmQ time intervals:
Fig. 9. Design flexibility exhibited for supply and return ducts.
ess ensures that the form instances selected are always unique in their geometry. The new forms evolve with changes in space (length, width, height), window (left distance, right distance, floor distance, ceiling distance) (Fig. 8); and supply inlet duct and return outlet duct (x distance, y distance, length, height) (Fig. 9). For every run of the GA algorithm, one unique form is generated. Therefore, morphing does not exist during the first cycle. For subsequent cycles, the morphing process runs continuously with a fixed time interval. Since the morphing is a direct linear transformation between
Fiþ1 YFi ½through bmQ time intervals: The forms used by the module represent threedimensional orthogonal spaces. The morphing proc-
Fig. 8. Design flexibility exhibited for space and window forms.
Fig. 10. Layer 4—design evaluation.
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44
41
Therefore, any in-between user-selected shape (denoted as F mi ) during the morphing process is again a set of geometric design variables: Fmi representsf RLmi ; RWmi ; RHmi ; W Lmi ; W Rmi ; W Fmi ; W Cmi ; SXmi ; SYmi ; SLmi ; SHmi ; RXmi ; RYmi ; RLmi ; RHmi g
Fig. 11. Design specification GUI.
where i=1, 2, 3,. . .n. two consecutive forms, any in-between shape can be identified as: f½Fiþ1 FFi =mg fselected time frame as a ratio of dmTg
where bmQ is the time interval in milliseconds.
2.4. Design Evaluation The Design Evaluation layer allows user intervention to take place. The intervention begins by the user selecting an instance from the evolving shape. Moving uninterrupted, the GA
Fig. 12. Morph-process GUI.
42
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44
continues to perform its search for the best designs. The selected shape can then be evaluated for its performance (Fig. 10). Performance can be evaluated by invoking the CFD engine. A mesh is automatically generated and sent with the corresponding boundary conditions for the CFD engine to be executed. The CFD outcome of temperature and velocity contours can then automatically be generated and channeled back to the Graphic User Interface (GUI).
3. System interface The system interface was developed using Java API and C++. The interface uses function calls to interact with the GA module, Gambit v2.0.4 and Fluent v6.0. Several GUI mechanisms were built to allow better control of the system (Fig. 11). These GUIs aid in data entry, modification and visualization in different modes. Data entry and modifications are
used to manipulate form and its properties, performance targets and constraints. Visualization aids the user to explore the design evolution, its form generation and performance outcome. The interface allows the user to select their mode of visualizing the forms, either as wire frame or shaded models. Once the CFD converges to a solution, the morph GUI is displayed. This GUI is composed of three sections. The two sections at either corner of the GUI represent the two form instances ( F i+1 and F i ) and their performances, while displaying the continuous morphing process in the center ( F mi ) (Fig. 12). During the morph process, the user can zoom, rotate and scale the form instances dynamically for better visual perception. The performance of the form instances can be visualized directly by selecting the iso-surface image. Five equidistant iso-surfaces along each axis are provided. Depending on the user selection, the corresponding CFD outcome of temperature and velocity contours are displayed enabling the user to
Fig. 13. A typical display of temperature contour.
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44
visualize the performance of the instance being investigated (Fig. 13). In general, the optimality of the final solution is difficult to prove for large problems with stochastic techniques such as genetic algorithms. However, within the current system, the user specifies a point of interest, and the algorithm looks for the optimum around that point. Hence, the user is provided with a solution that performs bbestQ within the user-specified area of the design space.
4. Discussions and conclusions The model described demonstrates the potential of exploring and visualizing the design evolution and its form generation based on a set of performance targets. In addition, this model was able to capture all possible design evolutions, provide a larger search space for the designers to interact with, and enable the user to visualize the performance of any design instance selected during morphing. By integrating an optimization module that creates discrete instances of design with a morphing module that captures these consecutive instances and morphs the instances in between, the system was able to provide an example of continuous evolution of optimization. Automation of the form generation based on a set of performance targets allows several designs to be evaluated in less computational time than typically performed by conventional means. When using complex simulations to evaluate designs, the relationship between the design variables and their corresponding performances are difficult to track. Thus, it is hard to make meaningful design moves manually by parametric studies. The GAs allow the designer to explore the design space and review the tradeoff associated with different design solutions in a systematic and rigorous manner. The automated process also has the advantage of not being biased by the designer’s view and hence has the potential to generate novel configurations that could otherwise remain unknown. The morph component on the other hand, promises solution(s), which may still not be the boptimumQ but takes into consideration the designer preferences. It allows the user to intervene in the search process and select designs
43
based on the designer preferences. These design instances represent points in the design space that are close to the GA selected instances, but may not be the output of the algorithm because their performance is similar to the one selected by the GA. In this manner, the user has active control over selecting and evaluating designs during the design generation process. One of the primary objective of performancebased design is to optimize the performance targets specified by the user—thermal and ventilation criteria, in this case. The effectiveness of a particular design to satisfy the performance targets measures the value of that design. Thus, if the design satisfies both requirements effectively, it could be referred to as a good solution. Although these discrete design instances are considered to be good solutions, it is difficult to ascertain that any intermediate instance during morphing would be a good solution that maintains the performance targets unless it is analyzed by the performance module. Model scalability for both problem description and performance evaluation was taken into consideration when building the model. The model has the ability to be expanded to enable the intermediate instances to be evaluated automatically. Presently, orthogonal geometries with typical vertical glass windows can be studied using this model. In future, the problem can be modified to include non-orthogonal geometries and other thermal performances. This will allow the proposed model better flexibility in design shapes. These modifications will allow more complicated forms to be generated. In addition, it will allow the possibilities of evolving a wide range of design performance scenarios such as tilt of glass, configuration of atriums, etc. This modification will also allow the model to be integrated with existing CAD/ CAM software. Such integration will facilitate both students and professional architectural designers to explore design spaces through evolution process, yet confirming to underlying building performance criteria.
Acknowledgements This work was funded by a grant from the University of Pennsylvania Research Foundation.
44
A.M. Malkawi et al. / Automation in Construction 14 (2005) 33–44
References [1] C. da Graca, Q. Chen, L.R. Glicksman, L.K. Norford, Simulation of wind-driven ventilative cooling systems for an apartment building in Beijing and Shanghai, Energy and Buildings 34 (2002) 1 – 11. [2] B. Eckhardt, L. Zori, Computer simulation helps keep down costs for NASA’s blifeboatQ for the international space station, Aircraft Engineering and Aerospace Technology, An International Journal 74 (2002) 442 – 446. [3] M. Bartak, I. Beausoleil-Morrison, J.A. Clarke, J. Denev, F. Drkal, M. Lain, I.A. Macdonald, A. Melikov, Z. Popiolek, P. Stankov, Integrating CFD and building simulation, Building and Environment 37 (2002) 865 – 871. [4] I. Beausoleil-Morrison, The adaptive conflation of computational fluid dynamics with whole-building thermal simulation, Energy and Buildings 34 (2002) 857 – 871. [5] Z. Zhai, Q. Chen, Solution characters of iterative coupling between energy simulation and CFD programs, Energy and Buildings 35 (2003) 493 – 505. [6] A.D. Quinn, M. Wilson, A.M. Reynolds, S.B. Couling, R.P. Hoxey, Modelling the dispersion of aerial pollutants from agricultural buildings—an evaluation of computational fluid dynamics (CFD), Computers and Electronics in Agriculture 30 (2001) 219 – 235. [7] S. Murakami, S. Kato, K. Ito, Q. Zhu, Modeling and CFD prediction for diffusion and adsorption within room with various adsorption isotherms, Indoor Air 13 (2003) 20 – 27. [8] H. Huang, F. Haghighat, Modelling of volatile organic compounds emission from dry building materials, Building and Environment 37 (2002) 1349 – 1360. [9] C. Topp, P.V. Nielsen, P. Heiselberg, Influence of local airflow on the pollutant emission from indoor building surfaces, Indoor Air 11 (2001) 162 – 170. [10] S.M. Lo, K.K. Yuen, W.Z. Lu, D.H. Chen, A CFD study of buoyancy effects on smoke spread in a refuge floor of a highrise building, Journal of Fire Sciences 20 (2002) 439 – 463. [11] H.B. Awbi, M.M. Nemri, Scale effect in room airflow studies, Energy and Buildings 14 (1990) 207 – 210. [12] Z. Jiang, Q. Chen, A. Moser, Comparison of displacement and mixing diffusers, Indoor Air 2 (1992) 168 – 179. [13] F. Haghighat, Y. Huo, J. Zhang, C.Y. Shaw, The influence of office furniture, workstation layouts, diffuser types and location on indoor air quality and thermal comfort at workstations, Indoor Air 6 (1996) 188 – 203. [14] J.J. Costa, L.A. Oliveira, D. Blay, Turbulent airflow in a room with a two-jet heating-ventilation system—a numerical parametric study, Energy and Buildings 32 (2000) 327 – 343. [15] H. Xing, A. Hatton, H.B. Awbi, A study of the air quality in the breathing zone in a room with displacement ventilation, Building and Environment 36 (2001) 809 – 820. [16] J. Mathews, M.Y. Rafiq, Adaptive search to assist in the conceptual design of concrete buildings, Developments in Neural Networks and Evolutionary Computing for Civil and Structural Engineering (1995) 179 – 187. [17] I.C. Parmee, M.A. Beck, An evolutionary, agent-assisted strategy for conceptual design space decomposition, Proceed-
[18]
[19]
[20]
[21]
[22] [23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33] [34] [35]
ings of the Evolutionary Computing, AISB International Workshop, Selected Papers, 1997, pp. 275 – 286. J. Kotani, M. Haqiwara, An evolutionary design-supportsystem with structural representation, Proceedings of the 26th Annual Conference of the IEEE Industrial Electronics Society, IECON 2000IEEE International Conference on Industrial Electronics, Control and Instrumentation, 2000, pp. 672 – 677. P.Y. Papalambros, D.J. Wilde, Principles of Optimal Design: Modeling and Computation, Cambridge Univ. Press, London, 2000. C. Soddu, Argenic design, European Academy of Design Contextual Design, Design in Context Conference, Stockholm, Sweden, 1997. S. Chien, Sheng-Fen, U. Flemming, Design space navigation in generative design systems, Automation in Construction 11 (2002) 1 – 22. S. Todd, W. Latham, Evolutionary Art and Computers, Academic Press, London, 1992. J.H. Jo, J.S. Geor, Space layout planning using an evolutionary approach, Artificial Intelligence in Engineering 12 (1998) 149 – 162. N. Bouchlaghem, Optimizing the design of building envelopes for thermal performance, Automation in Construction 10 (2000) 101 – 112. K.P. Lam, A. Mahdavi, S. Gupta, N.H. Wong, R. Brahme, Z. Kang, Integrated and distributed computational support for building performance evaluation, Advances in Engineering Software 3 (2002) 199 – 206. J. Michalek, R. Choudhary, P. Papalambros, Architectural layout design optimization, Engineering Optimization 34 (2002) 461 – 484. R. Pelletret, W. Keilholz, Coupling CAD tools and building simulation evaluators, Proceedings of the Sixth International Conference, International Building Performance Simulation Association, Kyoto, Japan, 1999, pp. 1197 – 1202. G. Suter, A. Mahdavi, Elements of a representation framework for performance-based design, Automation in Construction 39 (2004) 969 – 988. J.A. Wright, H.A. Loosemore, R. Farmani, Optimization of building thermal design and control by multi-criterion genetic algorithm, Energy and Buildings 34 (2002) 959 – 972. R. Choudhary, A. Malkawi, Integration of CFD and genetic algorithms, Proceedings of the Eighth International Conference on Air Distribution in Rooms, Copenhagen, Denmark, 2002. L.G. Caldas, L.K. Norford, A design optimization tool based on a genetic algorithm, Automation in Construction 11 (2002) 173 – 184. D.A. Coley, S. Schukat, Low-energy design: combining computer-based optimization and human judgment, Building and Environment 37 (2002) 1241 – 1247. M. Mitchell, An introduction to genetic algorithm, Complex Adaptive Systems Series, MIT Press, 1996. D. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley, Reading, MA, 1989. H.B. Awbi, Ventilation of Buildings, Chapman and Hall, London, 1991.
