1. INTRODUCTION

Designing is one of the most significant of human acts. It is one of the bases for change in our society. However, designers are amongst the least recognised of society's change agents. Surprisingly, given that designing has been occurring for many millennia, our understanding of the processes of designing is remarkably limited. Part of our understanding of designing comes not only from studying human designers as they design but from postulating design methods which describe some aspect of the design process without claiming to model the processes used by human designers. The early approaches to design methods were prescriptive when applied to human designers. More recently, design methods have been formalised not as humano-centred processes but as processes capable of computer implementation. Amongst the goals of these endeavours are to develop a better understanding of the processes of designing, to develop methods which can be computerised and to aid human designers through the introduction of novel methods which have no human counterpart. Much of this research is driven by the fact that human designs are very often incomplete, inadequate or just plainly poorly conceived for the task they are meant to address.

In this paper we include a brief description of some of the current research projects at the Key Centre of Design Computing that are indicative of the scope and content of our Computer Aided Architecutral Design (CAAD) research. The list of references at the end of the paper provides a more complete view of our research projects. The projects can be considered as addressing two major paradigms for CAAD research programs

(i) computational approaches, and

(ii) cognitive approaches.

The computational approach to CAAD research considers the role and application of computer technology in understanding design. There are two streams within the computational approach: computational models of design and computational support for designers. The cognitive approach considers formal models of human designers, with the ultimate goal of identifying better computational suppport for human designers. These programs provide a coherent umbrella under which the individual research projects fit and proceed.

2. COMPUTATIONAL MODELS

Computational models for CAAD comprise the development of models that effectively perform some aspect of the design process and models that assist the human designer perform some aspect of the design process. These two approaches could be referred to as design automation and design aid in the extremes, but typical projects include some aspect of both. We distinguish the approaches here to highlight the difference rather than to presume that such extremes are practical.

In the Key Centre of Design Computing we have several themes that fit within the computational models of design, including:

(i) case-based design

(ii) emergence in design

(iii) evolutionary systems in design

(iv) qualitative reasoning about space

Here we describe typical projects that use case-based reasoning and those that use evolutionary systems. The bibliography at the end of the paper provides references to other projects related to computational models of design.

There are three major themes under the computational support for designers approach:

(i) computer-mediated collaborative design

(ii) multimedia in design

(iii) virtual design studios

Here we will describe our experience with virtual design studios, again providing references to other projects in the bibliography.

2.1 Case-Based Design

Case-based reasoning (CBR) is a paradigm for problem solving which emphasises the role and use of situated experiences. In general, problem-solving using CBR is based on making analogies between the current problem-solving situation, and previously-encountered situationsstored in memory that provide information relevant to solving the new problem. The use of case-based reasoning as a process model of design involves the subtasks of recalling previously known designs from memory and adapting these design cases or subcases to fit the current design context. The development of this process model for a particular design domain requires a better understanding of how designs are or should be formally represented as well as how to formalise the process of recall and adapt. Two specific projects in which these aspects of case-based design research are considered are: CASECAD and CADSYN.

The objective of the CASECAD project is to propose and implement a representation for design cases that would support designers in finding a set of cases relevant to a design problem. CASECAD is a design browsing tool which allows a designer to either navigate through its design memory to find relevant design cases or to automatically retrieve design cases given a formal specifications of a new design problem. To pursue the objectives of the system we focused on a multimedia representation of cases so that the information would be readily understood and familiar to designers, and we focused on flexible memory indexing techniques so that a case could be retrieved using many different paths.

CASECAD uses an object-oriented multimedia representation for design cases. Each object includes a reference to CAD drawings and images, groups of attribute-value pairs, and text-based descriptions of the case. Cases are decomposed into subcases for ease of partial recall and the reuse of parts of a design. The attribute-value data is used for case organisation and recall, and the graphics and text descriptions are used to help the designer understand the case information. The attribute-value data is also helpful to designers, since it summarises important information in discrete packages, and provides an explicit interpretation of the meaning of the graphical representations.

The objective of the CADSYN project is to develop a representation of case memory that facilitates the development of models of design case adaptation. This focus on case adaptation was necessary in order to understand the types and the extent of the knowledge needed to allow case-based reasoning to be more than a memory-based system; that is, case-based reasoning could then provide a process model for generating "new" designs, not just recalling relevant ones.

The design adaptation process developed for CADSYN uses a propose-verify-modify cycle where the retrieved case is modified to match the new specifications - forming a proposed design solution. The verify-modify part of the cycle is modelled as a constraint satisfaction problem. Using this model for design adaptation, it is possible to distinguish between routine design, in which constraint satisfaction is possible with the given case memory, and non-routine design, in which the knowledge in case memory needs to be extended to find a feasible solution. The process model developed in the CADSYN project is shown in Figure 1.

2.2 Evolutionary Systems in Design

Evolutionary systems are computational systems loosely based on an analog with Darwinian evolution. They distinguish between a 'genetic' level representation and the resulting artifactual representation. They make use of genetic operators such as crossover and mutation. They work with populations of individuals rather than single individuals. The most common computational form of evolutionary systems are genetic algorithms. The most direct use of genetic algorithms in design is as a search process, searching for the most suitable values of design variables which achieve the best performances in the resulting design or designs. However, the projects here use genetic algorithms rather differently; they are used as a computational process to support other endeavours. Here we describe two specific projects: a co-evolutionary approach to design and the development of genetic engineering for learning design representations.

Co-evolutionary model of design

Evolutionary algorithms are relevant to design problem solving because they incorporate methods for generating design solutions with a focus for evaluating the fitness of the alternatives. A major focus in the use of Genetic Algorithms (GA) for design is optimisation. The assumption behind the optimisation perspective is that a fitness function is defined in advance. This assumption is not typical of the early stages of design where the design specifications are still changing. We develop a model of design in which the design specifications evolve in response to an evolving design solution. This model effectively defines an approach to co-evolutionary design.

The search space for design is usually ill-defined and is accompanied by ill-defined goals. Hence, part of a design process is to search for the definition of the problem. A model of design is presented, as illustrated in Figure 2, as the interaction of problem space and solution space. The problem space (or the functional requirements) is represented by P, and the solution space is represented by S. The co-evolutionary model of design has the following characteristics:

1. There are two distinct search spaces: Problem Space and Design Space.

2. These state spaces interact over a time spectrum.

3. Horizontal movement is an evolutionary process such that

a. Problem space P(t) evolves to P(t+1), P(t+2), and so on;

b. Solution space S(t) evolves to S(t+1), S(t+2), and so on.

4. Diagonal movement is a search process where goals lead to solution. This can be "Problem leads to Solution" (downward arrow) or "Solution refocusses the Problem" (upward arrow).

The problem space P(t) is the design goal at time t and S(t) is the solution space which defines the current search space for design solutions. The solution space S(t) provides not only a state space where a design solution can be found, but it also prompts new requirements for P(t+1) which were not in the original problem space, P(t). This is represented by the dashed upward arrow from design space S(t) to problem space P(t+1). The upward arrow is an inverse operation where S(t) becomes the goal and a "search" is carried out in the problem space, P(t+1), for a " solution". This iterative relationship between problem space and design space evolves over time.

This model of exploration depicts an evolutionary system, or in fact, two evolutionary systems. The evolutionary systems are the problem space and the solution space. The evolution of each space is guided by the most recent population in the other space. We have developed this model of design as a co-evolutionary system through the definition and implementation of two algorithms: CoGA1 and CoGA2, which make different assumptions regarding the representation of design genes and the implications of a global vs local fitness function.

Genetic engineering:

Genetic algorithms evolve designs from a fixed set of design 'genes' which go make up the 'genotype', genetic engineering is the process by which the genes which make up the representation of a design are allowed to evolve. It is based on the practice of genetic engineering in natural genetics where the genetic engineer intervenes in natural evolution by classifying that part of the population of the organisms available into one group with possesses the high level of the performance of interest and into another group that part of the population which does not. Then the genetic engineer tries to single out the groups of genes which are responsible for the high levels of the performance of interest (whose presence distinguishes this group from any other), Figure 3. Finally, the genetic engineer attempts to produce the next generation of the population using genotypes which contain more of these useful groups of genes utilizing various types of intervention in the reproduction process.

The concept can be illustrated using a simple 2-dimensional graphical example. Consider the eight assembly rules shown in Figure 4. These rules operate on a square shape. Any design can be coded as a sequence of these rules which are used to assemble it. Assume we are trying to produce designs with a maximum number of holes in it using only 20 squares. We start the process by producing a set of designs as shown in Figure 5. We evaluate these designs and find that four of them (designs 1, 2, 4 and 7) are good designs in that they have a high performance in terms of the number of holes.

Three of these four designs contain a gene group {2,8,5}. Further, none of the other designs contain this gene group. Thus, the next generation can use the existing 8 genes plus a new evolved gene which represents this gene group. The process can be repeated to evolve further genes which embody the desirable features. The effect of this process is an increasingly targeted representation ­ targeted towards the problem or class of problem being solved. Genetically engineered genetic algorithms have been used to solve problems more efficiently and as the basis of developing representations in case-based design.

2.3 Virtual Design Studios

The idea of a Virtual Design Studio (VDS) refers to a team of designers from various locations for which communication is computer-mediated; essentially, the studio is distributed across space and time and information is represented electronically. A Virtual Design Studio allows designers who are geographically dispersed to interact through their desktop computer. The location of the designers becomes irrelevant because the virtual studio is an electronically distributed workspace. Designers are able to enter the virtual studio that forms the common room in their interaction session by connecting to the World Wide Web and/or connecting to a video conferencing session. The information available to the participants in the design studio is stored in a variety of files and formats. As part of this project, we established a Virtual Design Studio as a collaborative project in a collection of Architecture Faculties. The student work resulting from these studios can be seen at: http://www.arch.su.edu.au/kcdc/

Real design projects require a collaboration of individuals and a coordination of information and tasks. Computer support for design, more specifically CAD systems, have been developed to support a single user through a graphical interface and project teams through file transfer and distributed data. Recent developments in networked communications has extended collaboration to include electronic mail and sharing information on the World Wide Web. Although these tools are used to varying degrees in professional practice, they have not yet reached their potential in supporting design processes.

When considering how computers can support design collaboration, two modes of collaboration can be identified: asynchronous and synchronous. In the asynchronous mode, designers may work at different times, often on different parts of the design, and do not require the simultaneous presence of all team members. Networked workstations have a wide variety of functions and tools that support sharing and exchanging information asynchronously, for example, Email, ftp, DBMS, etc. This type of collaboration is typical among design professionals in which the collaboration occurs across a long distance.

Synchronous collaboration implies the simultaneous presence and participation of all users, made possible by sharing applications among distributed designers, for example real-time conferencing systems, as illustrated in Figure 6, or shared drawing programs. While the current technology offers support for synchronous collaboration, the speed of the computers and the network traffic restrict the development of group interaction. This type of collaboration is less common in professional practice but shows promise as the technology improves.

The project on virtual design studios has resulted in a collection of examples of collaboratively designed projects represented on the WWW, a set of experiments in which professional designers use video conferencing to develop a design concept, and a number of system architectures that can support both synchronous and asynchronous design.

3. COGNITIVE APPROACHES

The research focus of the projects that fit within the cognitive approaches look closely at human designers for insight into effective computational models of design. The three major projects under the cognitive approaches are:

(i) design fixation

(ii) models of designing

(iii) role of diagrams in designing.

Here we will describe only the projects related to design fixation and models of designing.

3.1 Design Fixation

Designers are regularly exposed to example of previous designs for the same or similar design requirements. This work investigates the implicatations of this. Do designers unconsciously use features of designs they have been exposed to ­ if they do they exhibit a cognitive phenomenon called fixation. Figure 7 shows a drawing of a design for a device to aid the physically impaired in having a bath. Mechanical engineering designers and industrial designers were asked to design a device to aid the physically impaired in getting into and out of a bath. The control group was given only a verbal description of the requirements. The other group was given both the verbal description and a drawing of a solution, Figure 7, to indicate the degree of detail expected of them in their own documentation.

Figure 8 shows the preliminary results of one set of design fixation experiments. These indicate that mechanical engineering designers exhibit positive design fixation, ie they used proportionally many more features shown in the drawing than the control group. The industrial designers exhibited negative design fixation, ie they exhibited proportionally fewer features shown in the drawing than the control group.

Results such as these have significant implications for both the kinds of support to be provided to designers and to issues in the education of designers.

3.2 Models of Designing

Research is being carried out into design protocol methodologies and the data which can be derived from them to develop and support models of designing. A rich design-specific protocol, using the 'talk-aloud' method, has been developed which produces highly articulated results.

Figure 9 shows an example of some of the information which can be derived from such protocol analyses. Here the percentage of time, averaged over 10 minutes, devoted by a designer to function and behaviour is plotted against the time in the design session. This figure clearly shows that this designer has commenced by proposing a solution at the outset of the design session. This supports the hypothesis that certain designers design using a 'bottom-up' approach.

Figure 10 shows that this designer spends the majority of his time on events of less than 20 seconds duration.

Results such as these have significant implications for the computer-based tools needed to suppport designers. As we obtain more knowledge about various aspects of how humans design, so we are more able to provide tools which are likely to be more useful as design aids.

Acknowledgments

The work described in this paper has been funded by a number of Australian Research Council grants to the authors. The following research students and research staff have contributed to the research outlined here: Brett Boardman, Myung-Yeol Cha, Anna Cicognani, Jose Damski, Lan Ding, Andres Gomez, Jun Jo, Han Jun, Vladimir Kazakov, Tom McNeill, Soo-Hoon Park, Josiah Poon, Terry Purcell, Michael Rosenman, Thorsten Schnier, Simeon Simoff, Phil Tomlinson, Min Yan.

Bibliography of Recent Research Papers from the Key Centre of Design Computing

Alem, L. and Maher, M.L. (1994) A model of creative design using a genetic metaphor, in T. Dartnall (ed.) Artificial Intelligence and Creativity: An Interdisciplinary Approach, Kluwer, pp 281-291.

Damski, J. and Gero, J. S. (1994) Visual reasoning as visual re-interpretation through re -representation, AID'94 Workshop on Reasoning with Shapes in Design, Lausanne, pp.16-20.

Damski, J. and Gero, J. S. (1994) A model of shape emergence based on human perception, Information Technology in Design Vol. 2, ICSTI, Moscow, pp. 96-105.

Damski, J. C. and Gero, J. S. (1996) Shape reasoning in CAD systems, IEA-AIE'96 , Gordon and Breach, Langhorne, PA (to appear)

Damski, J. C. and Gero, J. S. (1996) A logic-based framework for shape representation, Computer-Aided Design 28(3):169-181.

Gero, J. S. (1994) Design and creativity, in T. Dartnall (ed.), AI and Creativity, Kluwer, Dordrecht, pp. 259-267.

Gero, J. S. (1994) Computational models of creative design process, in T. Dartnall (ed.), AI and Creativity , Kluwer, Dordrecht, pp. 269-281.

Gero J. S. (1994) Engineering design, computers and creativity: some explorations and prognosis, in D. Rehak (ed.), Bridging the Generations ­ Future Directions of Computer-Aided Engineering, Department of Civil Engineering, Carnegie Mellon University, Pittsburgh. pp.195 -200.

Gero, J. S. (1994) Exploration as a basis of creative engineering design, in K. Khozeimeh (ed.), Computing in Civil Engineering , ASCE, New York, pp.1825-1831.

Gero, J. S. (1995) Exploration, redescription and design creativity, Fourth International Workshop on Research Directions for Artificial Intelligence in Design, Enschede, pp. 35­40.

Gero, J. S. (1995) The role of function-behavior-structure models in design, in J. P. Mohsen (ed.), 2nd Congress on Computing in Civil Engineering, ASCE, New York, pp.294­304.

Gero, J. S. (1995) Recent advances in computational models of creative design, in P. Pahl and H. Werner (eds), Computing in Civil and Building Engineering, Balkema, Rotterdam. pp. 21-30.

Gero, J. S. (1995) Improving design problem formulations using machine learning, in J. P. Mohsen (ed.), 2nd Congress on Computing in Civil Engineering, ASCE, New York, pp.525­529.

Gero, J. S. (1995) Computers and creative design, CAAD Futures '95, National University of Singapore, pp.3.1­9.

Gero, J. S. (1995) Computers in creative design, ARCHCOMP, SAFA, Helsinki, pp.1­13.

Gero, J. S. (1996) Creativity, emergence and evolution in design: concepts and framework, Knowledge-Based Systems (to appear)

Gero, J. S. (1996) Advances in formal design methods for computer-aided design, in Gero, J. S. (ed.) Advances in Formal Design Methods for CAD , Chapman and Hall, London, pp. 293-298.

Gero, J. S. and Damski, J. (1994) Object emergence in 3D using a data driven approach, in J. S. Gero and F. S. Sudweeks (eds), Artificial Intelligence in Design '94 , Kluwer, Dordrecht, pp. 419-436.

Gero, J. S., Damski, J. and Jun, H. (1995) Emergent shape semantics, CAAD Futures '95, , National University of Singapore, pp.49.1­15.

Gero, J. S. and Jun, H. (1995) Visual semantic emergence to support creative design: a computational view, Preprints Computational Models of Creative Design , University of Sydney, pp.87­116.

Gero, J. S. and Jun, H. (1995) Getting computers to read the architectural semantics of drawings, in L. Kalisperis and B. Kolarevic (eds), Computing in Design: Enabling, Capturing and Sharing Ideas, ACADIA, Washington. pp.97-112.

Gero, J. S. and Kazakov, V. (1996) An exploration-based evolutionary model of a generative design process, Microcomputers in Civil Engineering 11:209-216.

Gero, J. S. and Kazakov. V. (1996) Evolving building blocks for design using genetic engiineering: a formal approach, in J. S. Gero (ed.), Advances in Formal Design Methods for CAD, Chapman and Hall, London, pp.31-50.

Gero, J. S., Kazakov, V and Schnier, T. (1996) Evolving design genes as well as design solutions, Third International Congress on Computing in Civil Engineering, ASCE, New York (to appear)

Gero, J.S., Louis, S. and Kundu, S. (1994) Evolutionary learning of novel grammars for design improvement, AIEDAM 8(2):83-94.

Gero, J. S. and Louis, S. (1995) Improving Pareto optimal designs using an evolutionary approach, Microcomputers in Civil Engineering 10 (4): 241­249.

Gero, J. S. and Schnier, T. (1995) Evolving representations of design cases and their use in creative design, Preprints Computational Models of Creative Design , University of Sydney, pp.343­368.

Gero, J. S. and Yan, M. (1994) Shape emergence using symbolic reasoning, Environment and Planning B: Planning and Design 21: 191-212.

Gomez de Silva Garza, A. and Maher, M.L. (1996). Design by in teractive exploration using memory -based techniques, Knowledge-Based Systems, 9(1).

Jo, J. and Gero, J. S. (1995) Representation and use of design knowledge in evolutionary design, CAAD Futures '95, , National University of Singapore, pp.17.1­16.

Jo, J. and Gero, J. S. (1995) A genetic search approach to space layout planning, Architectural Science Review 38 (1): 37­46.

Jo, J. and Gero, J. S. (1996) Space layout planning using an evolutionary approach, Artificial Intelligence in Engineering (to appear)

Maher, M.L. (1994). Representation of case memory for structural design, Computing in Civil Engineering, ASCE, 2030-2037.

Maher, M.L. (1994) Creative design using a genetic algorithm, Computing in Civil Engineering, ASCE, pp 2014-2021.

Maher, M.L. (1994) Collaboration and computer aided engineering, Bridging the Generations: An International Workshop on the Future Directions of Computer-Aided Engineering , Carnegie Mellon University, Pittsburgh, PA, pp137-142.

Maher, M.L., Balachandran, B., Zhang, D.M. (1995) Case-Based Reasoning in Design, Lawrence Erlbaum Associates.

Maher, M.L. (1995) Using case-based reasoning for design media management, Computing in Civil Engineering, ASCE, Atlanta.

Maher, M.L. (1995) Using the internet to teach in a virtual design studio, DECA 95: Information Technology and Its Influence on Design Education, RMIT, Melbourne.

Maher, M.L. (1995) An experimental study of computer-mediated collaborative design, Design Challenges, Queensland University of Technology.

Maher, M.L. (1996) CASECAD and CADSYN: Implementing case retrieval and case adaptation, in M.L. Maher and P. Pu (eds) Issues and Applications of Case-Based Reasoning in Design , Lawrence Erlbaum.

Maher, M.L. and Balachandran, B. (1994) A multimedia approach to case-based structural design, ASCE Journal of Computing in Civil Engineering , 8(3): 359-376.

Maher, M.L. and Balachandran, B. (1994) Flexible retrieval strategies for case-based design, Artificial Intelligence in Design'94, J. Gero (ed.), Kluwer Academic Press, pp 163-180.

Maher, M.L., Boulanger, S., Poon, J., and Gomez de Silva Garza, A. (1995) Assessing computational methods with a framework for creative design processes, Preprints Computational Models of Creative Design, University of Sydney, pp.233-265.

Maher, M.L. and Gomez de Silva Garza, A. (1996) Developing case-based reasoning for structural design, IEEE Expert, (to appear).

Maher, M.L. and Harwood, B. (1994) Design media management, Proceedings of Multimedia and Design Conference , Key Centre of Design Computing, University of Sydney, pp 219-232.

Maher, M.L. and Poon, J. (1996) Modelling design exploration as co-evolution, Microcomputers in Civil Engineering, 11:192-207.

Maher, J. and Poon, J. (1995) Evolving a fitness landscape for design exploraiton, International Conference on Evolutionary Computing , Perth, Australia.

Maher, M.L., Poon, J. and Boulanger, S. (1995). Formalising Design Exploration as Co-Evolution: A Combined Gene Approach, in J.S.Gero and F. Sudweeks (eds) Preprints Formal Design Methods for CAD, pp 1-28.

Maher, M. L. and Pu, P. (eds) (1996) Issues and Applications of Case-Based Reasoning to Design, Lawrence Erlbaum Associates, in progress.

Maher, M.L. and Rutherford, J. (1996) A model for collaborative design using CAD and database management, Research in Engineering Design,(to appear).

Maher, M.L. and Saad, M. (1995) The experience of virtual design studios at The University of Sydney, 1995 ANZAScA Conference, University of Canberra.

Purcell, A. T., Gero, J. S., Edwards, H. and Matka, E. (1994) Design fixation and intelligent design assistants, in J. S. Gero and F. S. Sudweeks, (eds) Artificial Intelligence in Design '94, Kluwer, Dordrecht, pp. 483-496.

Purcell, A. T., Gero, J. S., Edwards, H. and McNeill, T. (1996) The data in design protocols: The issue of data coding, data analysis in the development of models of the design process, in K. Dorst, H. Christiaans and N. Cross (eds), Analysing Design Activity, John Wiley, Chichester (to appear).

Qian, L. and Gero, J. S. (1996) Function-behaviour-structure paths and their role in analogy-based design, AIEDAM (to appear)

Qian, L. and Gero, J. S. (1995) An approach to design exploration using analogy, Preprints Computational Models of Creative Design , University of Sydney, pp.3­16.

Rosenman, M. A. and Gero, J. S. (1994) The what, the how and the why in design, Applications of Artificial Intelligence 8 (2):199-218.

Rosenman, M. A. and Gero, J. S. (1996) Modelling multiple views of design objects in a collaborative CAD environment, Computer-Aided Design 28(3):193-205.

Rosenman, M., Gero, J.S., and Maher, M.L. (1994) Knowledge-based design research at the Key Centre of Design Computing, in G. Carrara and Y. Kalay (eds) Knowledge-Based Computer-Aided Architrectural Design , Elsevier Science. (also published in Automation in Construction 3:229-237.)

Schnier, T. and Gero, J. S. (1996) Learning genetic representations as alternative to hand-coded shape grammars, in J. S. Gero and F. Sudweeks (eds), Artificial Intelligence in Design'96, Kluwer, Dordrecht, pp.39-57.

Schnier, T. and Gero, J. S. (1995) Learning representations for evolutionary computation, in X. Yao (ed.) 8-AJC Artificial Intelligence, World Scientific, Singapore. pp. 387-394.

Saad, M. and Maher, M.L. (1994) Multimedia and synchronous collaborative design, Proceedings of Multimedia and Design Conference , Key Centre of Design Computing, University of Sydney, pp 103-119.

Saad, M. and Maher, M.L. (1995) Exploring the possibilities for computer-supported collaborative designing, CAAD Futures'95, Singapore.

Saad, M. and Maher, M.L. (1996) Shared understanding in computer-supported collaborative design, Computer-Aided Design 28(3):183-192.

1 The Key Centre of Design Computing, originally set up with direct funding from the Australian Federal Government under its key centres program, built upon a pre-existing centre: the Design Computing Unit. The Key Centre currently has around 20 teaching, research and support staff and 15 doctoral students in the area of design computing in addition to some 60 graduate students taking coursework degrees.