Useful Mechanisms for Developing Simulations for Cognitive Models
Contents
1. Recommendations We end up with two recommendations The first which you could do now is to use the existing software and systems recommended above The analyst will prefer to use or extend an existing simulation if one is available that meets their criteria In many cases it appears that simulations have not been built with cognitive modelling in mind so this will not be possible and a simulation will have to be built There is a range of simulation building tools available starting with Soar Sim which is adequate for simple models that simply process set inputs Slightly more complex simulations can be implemented in languages like cT which is easy to work with More complicated simulations requiring more powerful toolkits will find Garnet useful Most analysts will wish to develop the model using the a structured integrated editor which currently requires a Unix machine for Soar at least Sockets will be the communication medium of choice using the MONGSU library The second recommendation is to include an interpreted extension language Garnet provides this When complete TclSoar will show that the Tcl can serve as well We have started developing both the ATC task Ong amp Ritter 1995 and the Tower of Nottingham task with Garnet The models and simulations run on a Unix machine communicating using sockets The subjects version of the ATC simulation runs on the Macintosh We will not immediately create a version of the Tower of2. Kosbie D S Pervin E Mickish A amp Marchal P 1990 Garnet Comprehensive support for graphical highly interactive user interfaces IEEE Computer 23 11 71 85 Nelson G 1994 SimTime User s Guide FTPable from centro soar cs cmu edu afs cs project soar public Soar6 user library SimTime tar Z The Soar Group School of Computer Science Carnegie Mellon University Newell A 1990 Unified Theories of Cognition Cambridge MA Harvard University Press Nichols S amp Ritter F E in press A theoretically motivated tool for automatically generating command aliases In Proceedings of CHI 95 Ong R amp Ritter F E 1995 Mechanisms for routinely tying cognitive models to interactive simulations In HCI International 95 Osaka Japan Pearson D J Huffman S B Willis M B Laird J E amp Jones R M 1993 A symbolic solution to intelligent real time control Robotics and Autonomous Systems 11 279 291 Peck V A amp John B E 1992 Browser Soar A computational model of a highly interactive task In CHI 92 Conference on Human Factors and Computing Systems 165 172 New York ACM Press Regian J W amp Shute V J 1993 Basic research on the pedagogy of automated instruction In D M Towne T de Jong amp H Spada Eds Simulation based experiential learning pp 121 132 Berlin Springer Verlag Ritter F E Hucka M amp McGinnis T F 1992 Soar mode
3. Nottingham for running subjects because we believe that we have adequate data at this point and because validating the simulation as comparable to real blocks could be quite difficult References An archive of Soar software is available via anonymous FTP from host centro soar cs cmu edu in the directory afs cs cmu edu project soar public Soar6 If you start to use any of this code please email soar requests cs cmu edu to notify them so that they can add you to appropriate mailing lists provide updates and recognise you when you send in bug reports Anderson T 1994 May Visual Basic for applications Personal Computer World 492 496 Glass G 1993 Unix for programmers and users A complete guide Englewood Cliffs NJ Prentice Hall Guttman R H amp Huffman S B 1992 SoarSim A Soar simulation building tool version 1 1 user s manual Artificial Intelligence Laboratory U of Michigan Hucka M 1994 The Soar Development Environment Artificial Intelligence Laboratory University of Michigan Jones O 1989 Introduction to the X Window System Englewood Cliffs NJ Prentice Hall Lehman J F Newell A Newell P Altmann E Ritter F amp McGinnis T 1994 The Soar video The Soar Group McMullan J amp Steier D 1994 WinSoar Soar for IBM PC compatibles running MS Windows In T Johnson Ed Soar Workshop 13 164 165 Myers B A Giuse D A Dannenberg R B Vander Zanden V
4. an acknowledgement here does not indicate an endorsement Support has been provided by the DRA contract 2024 004 and by the ESRC Centre for Research on Development Instruction and Training Introduction A general need when creating cognitive models is providing a simulation for them to manipulate These models need a simulation to interact with and learn from and subject behaviour to build and test the model acquired from the same simulation or an equivalent Previous attempts have used a variety of strategies Some simulations have provided the model with a set of fixed inputs that require the model to do the same action as the subject did Peck amp John 1992 Others have implemented the task such as the Tower of Hanoi in the modelling language itself Ruiz amp Newell 1989 More ambitious work has provided the cognitive model with direct access to a simulation often but not always by working within an integrated environment like the Lisp Machine This has been done in the Tacair Soar system Rosenbloom Johnson Jones Koss Laird Lehman et al 1994 This is an advanced simulation both from the standpoint of including multiple aircraft the scope of the cognitive model and the technologies that have been developed to support the creation and integration of its parts But the scale of this effort perhaps 60 months of effort emphasises that providing this access is not yet routine Figure 1 shows the possible relationships b
5. and writes We do not recommend it but introduce it first as it is a precursor to Unix sockets which are similar but more robust Sockets Sockets are a well used Unix mechanism including AUX on the Macintosh for allowing two processes to communicate Further information is readily available in textbooks such as Glass 1993 They are similar to files in how processes read and write them but they do not have disk space allocated actually when a disk file is manipulated this is also done with a socket but this is transparent to the user Sockets are more costly to set up in programming time than we currently would like they are a relatively low level communications medium but relatively robust and portable and can run across and between machines They offer a relatively high process communication bandwidth Starting with code generated by Mertz and Pelton we have created a package MONGSU available in the Soar archive that makes using this channel simpler Ong amp Ritter 1995 While it was designed for Soar and Garnet any combination of C and Lisp systems that need to communicate may benefit GNU Emacs GNU Emacs runs on Unix VMS and DOS platforms There are copies emerging for the Mac but they are of no use here because they do not yet incorporate process control such as running a model Emacs offers a relatively simple lightweight approach to tying processes together When a process spawned within Emacs returns a value a
6. creates prize winning portable physics programs Computers in Physics 7 2 136 143 Stevens W R 1990 Unix Network Programming Englewood Cliffs NJ Prentice Hall Wood D Bruner J S amp Ross G 1976 The role of tutoring in problem solving Journal of Child Psychology and Psychiatry 17 89 100 Wood D Shadbolt N Reichgelt H Wood H amp Paskiewitz T 1992 EXPLAIN Experiments in planning and instruction AJSB Quarterly 81 13 16
7. set of patterns are used to assign an associated Lisp function to handle the output Typically this is simply to print out the text but this also offers a simple way to pass messages between two processes It could be used to tie the model quickly to a simulation This would offer several advantages if it were available as a utility which it might be possible to build fairly easily This system would be flexible because it is implemented in an interpreted language that includes a debugger It could be extended easily to pass information to several processes at once or to record messages for later analyses Experience with SDE Hucka 1994 suggests that the bandwidth of this approach is Good This approach has one main disadvantage This mechanism is less robust because it is possible for the user to interrupt the handler perhaps by mistake Microsoft Windows interprocess communication Dynamic Link Libraries DLL provide a means of sharing functions between different Microsoft Windows applications These functions have to be linked in but are fairly straightforward for those who know them They tend to be used for side effects and not for passing information Dynamic data exchange DDE is a system for automatically passing data between Windows applications DLL is similar to joint compilation and DDE similar to sockets Both of these essentially represent joint compilation in an environment designed to support this approach Microsoft
8. that provide pre emptive scheduling but these are rare and require replacing the standard operating system Of course with a PC or Mac there are no other process running so timing done right is less of an issue Logs can be created from videotaping the users but this is expensive So while machines running Unix are likely to be the machine of choice for developing models and simulations another machine running the simulation must be used to record the subject s data Micro computers Micro computers can provide accurate timing information If platform independent software is used to develop the simulation subjects can be run on both platforms The drawbacks that preclude the use of Macintoshes as a complete environment include the lack of model development environment and the relative difficulty of inter process communication PC equivalents 386 or better can provide acceptable environments for developing models They appear to support more stand alone simulations than the other hardware platforms combined What they lack is a superior model development environment a complete simulation development environment and a robust and easy to use communication medium We expect that the PowerMacintosh a cross between a Macintosh and a PC will become important for developing systems It is a faster and has a more open architecture Recent press releases from Apple and Digitool suggest that MacLisp will be ported to the PowerPC s native code
9. wrote both Windows and the major horizontal applications This would be a good way to interface Soar to existing software probably much easier than using Apple Events but not useful for Soar running under the UNIX environment Modelling perception of what is on the screen would still be difficult to add if applications did not include it Apple events Apple events are objects messages that can be sent to most programs running on the Macintosh These events correspond to user input such as keystrokes and mouse movements or pre specified queries about contents There are three problems with using this First while it is intrinsic to the operating system relatively less work has gone into developing reading events such as What s in location 2 3 Reading events require a separate function to be written and linked into the simulation This is not different from a socket handler but the technology of developing such programs appears to be less mature Second finding expertise in this area in people and documentation appears to be difficult making its use more difficult Third there are rumours that it will not be supported in later versions of the Macintosh operating systems Joint compilation linking The fastest and most robust way to link two processes together is to compile them together and have them call each other This approach is used to link Soar and the ModSAF airplane simulation environment Schwamb Koss amp Keirs
10. Manual CMU CS 92 205 School of Computer Science Carnegie Mellon University Ritter F E amp Larkin J H 1994 in press Using process models to summarize sequences of human actions Human Computer Interaction Ritter F E amp Young R M 1994 Practical introduction to the Soar cognitive architecture Tutorial report AJSB Quarterly 88 62 Rosenbloom P S Johnson W L Jones R M Koss F Laird J E Lehman J F Rubinoff R Schwamb K B amp Tambe M 1994 Intelligent automated agents for tactical air simulation A progress report In Proceedings of the Fourth Conference on computer generated forces and behavioral representation Gainsville U of Central Florida Rosenbloom P S Laird J E amp Newell A 1992 The Soar papers Research on integrated intelligence Cambridge MA MIT Press Ruiz D amp Newell A 1989 Tower noticing triggers strategy change in the Tower of Hanoi A Soar model In The Annual Conference of the Cognitive Science Society 522 529 Schwamb K 1994 TclSoar A programmable Soar shell In Soar Workshop XIV 73 79 U of Michigan Schwamb K B Koss F V amp Keirsey D 1994 Working with ModSAF Interfaces for programs and users In Proceedings of the Fourth Conference on Computer Generated Forces and Behavioral Representation Technical Report IST TR 94 12 Institute for Simulation and Training Sherwood B A amp Andersen D M 1993 cT
11. Ritter F E amp Major N P 1995 Useful mechanisms for developing simulations for cognitive models AJ and Simulation of Behaviour Quarterly 91 Spring 7 18 Useful Mechanisms for Developing Simulations for Cognitive Models Frank E Ritter Nigel P Major ESRC Centre for Research in Development Instruction and Training Department of Psychology University of Nottingham Nottingham NG7 2RD UK Email ritter nigel psyc nott ac uk Phone 44 115 951 5302 Fax 951 5324 Abstract We describe an approach for developing cognitive models that can interact with real time interactive tasks like flying a plane and static puzzle like tasks The primary requirement is the ability to create simulations usable by both subjects and models Additional general simulation requirements can be noted for our applications in particular such as visual displays We also consider desirable model environment features After a description of our sample tasks and a survey of near misses we defend our choices of a using Garnet as a simulation language b developing a Soar model under Unix and the Soar Development Environment and c the use of sockets as a way to hook them together Acknowledgements We thank Martin Arnold Ellen Bass Mike Hucka Howard Martin Joe Mertz Brad Myers Sarah Nichols Roberto Ong Karl Schwamb Bruce Sherwood and David Steier for comments and useful discussions Any errors remain the fault of the authors and
12. S 70 and runs on the PC or Macintosh New scenarios can be added but it cannot be otherwise modified instrumented or run by other software A complete version is available for approximately 25 000 that runs on dedicated hardware This version is used at the US Federal Aviation Agency Academy to train tower operators and for research SGI flight simulator Some simulators come bundled with the operating system or the hardware For example there is a flight simulator available that runs on and takes advantage of the special hardware features of Silicon Graphics computers While it is not completely appropriate for our simulation it is interesting because a Soar system has been created that flies it Pearson Huffman Willis Laird amp Jones 1993 General purpose simulation languages If an appropriate simulation is not available one must be built This can be done in a standard programming language but special simulation and interface languages and toolkits can ease the task of creating a simulation that will run on multiple machines We currently prefer the Garnet system and have started an effort to create a simple ATC simulation in it that will meet the requirements noted in Tables and 2 Ong amp Ritter 1995 The simulation will need to run on what is natural for the model this will be Unix for Soar and what is natural for the subject this will be a system that can provide accurate timing information such as the Macintosh o
13. acs and PCs It is to be as easy to learn as SUIT noted below but capable of scaling up to full size interfaces There will be some kind of interpreted language eventually similar to function to Tcl such as Scheme interpreted C or both to describe an object s behaviour rather than a new language cT cT Sherwood amp Andersen 1993 is a Pascal like language developed by people sympathetic to tutoring systems Modellers interested in simpler simulations and who don t already know Lisp would find this a particularly attractive language to work with It includes a visual display and simple user interface components such as menus Source code including dis plays can run on Unix machines Macs and PCs It has been used to provide several Soar models with simulations using sockets on Unix systems and can provide the same inter process communication ability pseudo sockets across networked Macintoshes cT becomes unsuitable for large simulations such as the ATC task because it is a relatively low level language cT supports multi dimensional arrays with bounds checking for example but no data or record structures There will soon be a version 2 5 with significant enhancements for multimedia such as support for QuickTime and Video for Windows The Unix version is freely available via anonymous FTP from columbus andrew cmu edu in the directory pub ct The DOS and Macintosh versions are available from Physics Academic Software PO Bo
14. at others may wish to consider Modellers may be able to find other simulations related to their needs Table 4 lists several of the advantages of using an existing simulation Despite these advantages few simulations meet the needs listed in Table 2 communication between the model and the simulation typically is a problem particularly extending the simulation to provide reasonable perception to the model Simulators at Brooks AFB Psychology labs studying learning or performance will often have existing tasks that they use For example there are several simulation development efforts going on at Armstrong Lab at Brooks Air Force Base to support investigations into individual differences intelligent tutoring systems and rapid decision making Regian amp Shute 1993 As part of their work on studying automated instruction the group has developed 14 criterion tasks A typical example is Space Fortress a well studied task see Vol 71 of Acta Psychologica 1989 It is essentially an asteroids like video game These simulations do not appear to be appropriate for the initial simulation of studying rapid decision making or children s development but are likely to be useful as an additional task for studying rapid decision making TRACON simulations Simulations may also be available commercially Wesson International of Austin TX ph 512 328 0100 offers a range of ATC simulators called TRACON The simplest version costs approximately U
15. ding the Tower of Nottingham a pyramid made out of 21 blocks the other is an attempt to model real time problem solving and the development of rapid decision making in a simple Air Traffic Control ATC task In each case it is not possible to provide the cognitive model in our case Soar Newell 1990 Ritter amp Young 1994 Rosenbloom Laird amp Newell 1992 with a simulation written solely in Soar productions or by providing a fixed set of inputs provided at set times These tasks have individual requirements noted below and several shared requirements shown in Table 1 that will be common across most tasks These requirements form a foundation for choosing an approach to creating simulations In many cases the simulation tasks are likely to evolve The simulations are not initially well specified and will develop over time Once the model has been created and initially tested it will also suggest modifications either to explore additional behaviour or as a way to explore the effect of the world on the task solution These sorts of changes can be anticipated for both of our example tasks The initial model of solving the Tower of Nottingham is likely to suggest additional features of the blocks and their relationships that are necessary to include as perceptual input we will start with a simplified view The simple Air Traffic Control like task is also not completely specified We foresee additional modification based on the quest
16. e models The distributed Soar system provides a command line interface and runs after appropriate compilation on Macintoshes PCs and Unix machines These versions are available from the archive noted in the references and tend to be named based on their release number Aliases Nichols amp Ritter in press are included in the latest Soar release 6 2 5 which offers some customisability There are large differences between the existing environments for developing Soar models The development environment for Soar under Unix is good enough to force the model to be developed there and thus the simulation Soar and the environments described here are all in the public domain so price does not affect the choice Integrated structured editors SDE Hucka 1994 is currently the best environment for manipulating Soar models As a menu driven editor integrated with the running Soar process it recognises productions and allows the modeller to manipulate them in a natural way doing such things as directly loading or tracing them It also includes some useful displays and commands for finding and structuring the model It is similar in many ways to the interface provided by Clips It is a set of modifications to GNU Emacs similar to and vaguely evolved from earlier efforts with a relatively long history Ritter Hucka amp McGinnis 1992 It is still being developed and users can modify or extend it as well It requires Emacs 19 23 and E
17. ed as a relatively static current display In our task we only expect to have the circular display which will be more dynamic With this task we are definitely not trying to duplicate ATC instead we are accepting a simplified version so that we can look at the following questions e How do people learn to do a real time process control task e What skills and background knowledge do they need to start the task Tn order to avoid confusion and for brevity we define the model to refer to the cognitive model and the simulation to refer to the simulation of the world that the model uses Table 1 Requirements for simulations to be used with cognitive modelling e What the subject sees should be visible to the model This implies a theory of visual processing that translates the simulation s objects and provides them to the model What the model sees should also be visible to the analyst e The simulation must be modifiable Initial versions of the simulation and of the model s perception may be simplistic but in these tasks we expect the visual format the task operations and their perception to play a larger role over time so the simu lation and its interface to the model must be expandable e The simulation s actions must be recorded A basic trace is required showing what the subject and model saw or could have seen A complete trace which would be awkward to use routinely must also be available for detailed analysis This n
18. eed not be stored although this is safer in case the simulation changes but may be derivable from a listing of key events e Subjects actions and their times must be recorded Most cognitive models do not provide predictions with more than 100 ms accuracy so the actions timestamps need not be more accurate than 100 ms e All controls must be manipulable by the model and subject In order to explore computer supported work as in Figure Ic it is necessary to allow subjects and the 17 to inter he simulati e It must include time stress requiring decisions to be made rapidly from several options This translates into 10 50 planes per 10 min scenario These numbers are an order of magnitude higher than real ATC but the amount of handling is relatively small No negotiation with the aircraft or other control sectors is expected e A visual display for subjects and analysts subjects so that they can solve the task and analysts so that they can understand what the model sees its behaviour and internal state e A dynamic environment must be provided Monitoring is an essential aspect of this task so an environment must be provided with changes not all model initiated e Perceptual features of the display must be modifiable and provided to the model The perceptual ease with which subjects can recognise objects and obtain information is likely to influence how the task is performed e The task should force subjects
19. el will consist solely of keystrokes and mouse movements and clicks although simulated verbal output may be added later The output will also be symbolised and we expect that it will not exceed 10 symbols per second We expect to use sockets in both of our actual tasks but may briefly explore the feasibility of using Emacs processes as well There are other methodologies and mechanisms for interprocess communication Stevens Table 5 Features of various communication channels Bandwidt Ease of Flexibilit Robustne h creation y ss a Plain files Low Good Bad OK b Sockets High OK OK Good c Emacs Medium OK Good Maybe bad processes d Joint Compilation V High Bad OK V Good e Apple events Medium Bad Bad OK 1990 provides more Unix possibilities We only offer the most likely candidates particularly ones that can run between and on multiple platforms or on the platforms we will have to use Plain file reading and writing The simplest and quickest method of tying two processes together is to have one write to a disk file and the other to read it This approach gets complicated when multiple messages are passed and the read write cycles become often enough that the two processes contend for the same file if they don t actually accidentally overwrite it although there are semaphore techniques available to avoid clashes This approach thus suffers from low bandwidth and complete reliance on the operating system to adjudicate reads
20. etween a cognitive model and a simulation In Figure la two simulations must be created one to collect data from subjects and one for the model to manipulate The simulation for the subjects might not be a simulation but a real world task realised in hardware instead of software This approach increases the required work and questions can be raised about the equivalence of the two simulations Figure 1b is a better situation where the model s and the subjects view of the task are equivalent limited only by theories of perception This approach is an important and necessary step if one is interested in creating the situation in Figure 1c where the cognitive model assists the subject We are interested in creating an environment that provides these later two situations We review here the various system components that we are assembling to support creating the situation in Figure 1b and 1c for models of two tasks which we will use as example tasks a b c For Interactive Embedded knowledge real time user model analysis model or assistant Subject Subject Subject J 4 T ae Interactive Interactive Bie world world Interactive conus world mode Cognitive A model Cognitive model Figure 1 Possible relationships between models and simulations Shared features of these and other tasks We have started to model human problem solving in two areas that involve extensive interaction One is a blocks world type task buil
21. ey 1994 The advantages of this approach are that whole structures can be passed as they are not actually separate processes and that the communication rate is as fast as possible We believe that the disadvantages of joint compilation will outweigh the advantages for most users The mutual calling of functions requires the simulation for Soar to be written in C and duplicates what are normally operating system functions Changes to one of the systems requires knowledge of the other s internal functions and data structures and relinking or recompiling Because there is no intermediary observing and thus debugging this link may be difficult Hardware platforms The requirements and suggested solutions in the previous sections for developing the cognitive model developing the simulation and the process communication channel to use directly influence the choice of computer These needs split into two set of requirements met by different machines We will need to use two systems one for developing the model and one for running subjects It is worth mentioning an additional timing problem that may occur Unix and other multi tasking systems have essentially non deterministic schedulers This is the time and order of running the model and the simulation will vary On large time scales there is plenty of time for the display information to be passed to the model and for the model to pass back commands On short time scales the model and si
22. ion of how to facilitate the task either through modifying the display or through the basic support the display provides Figure 2a shows how the tower can be built out of blocks that fit together to create a ziggurat stepped pyramid Figure 2b shows how two pieces fit together to create a half level Two half levels make a level and five levels make the tower along with a capping piece An extensive set of regularities of childrens behaviour is available including while they were taught how to solve it by teachers Wood Bruner amp Ross 1976 other children and a computer tutor Wood Shadbolt Reichgelt Wood amp Paskiewitz 1992 This task makes it clear why the analyst must also see what the model and subject see including a visual display of what the model could see and what it actually attends to The position and orientation of the blocks are crucial in performing this task This information is complicated and the model s behaviour could vary wildly with small changes We are also interested in studying the genesis of rapid decision making Ong amp Ritter 1995 and have chosen to start with a simple ATC like task Because we are free to make changes to suit our needs without regard to external validity our task will differ from real ATC in numerous ways although some of the results are likely to transfer For example real ATC uses paper flight strips to keep track of planes and the big circular displays one expects are us
23. is a C based system developed at the U of Virginia It promises to be easy to use up and running simple applications in 2 hours estimated time It is freely available via FTP and is actively maintained If the Garnet based system is scrapped we are likely to take this system very seriously For more information finger suit uvacs cs virginia edu Winterp WINTERP is a object oriented toolkit developed in XLISP a public domain Lisp It provides a number of advanced features coupled with the flexibility and power of a small efficient Lisp language and object system It is available via anonymous FTP at ftp x org in contrib devel_tools Model simulation communication channels As the model runs it must query the simulation for input and provide the simulation with task commands Table 5 provides a summary of the features of various communication channels for tying models to simulations For us a communication medium s ease of use flexibility and robustness under use is more important than its bandwidth The amount of information in tasks pass between models and their simulations can be provided by most communication media At present the input from the simulation solely represents the result of visual scanning of the simulation s displays but we anticipate incorporating a verbal input as well When presented symbolically we expect that this will represent no more than 100 symbols per second At present the output from the mod
24. macs ability to manipulate processes which is currently available only under the Unix or Linux operating systems SDE is available from the centro archive An interface to Soar based under Microsoft Windows 3 x has also been developed McMullan amp Steier 1994 It includes a production editor dialogue boxes for setting various parameters context sensitive help and several other features However it is not as complete an environment as the SDE and its top speed on the fastest PCs does not match the speed of the fastest Unix machines WinSoar is made available to the Soar community on an unsupported basis but users with only a PC should find it quite valuable and adequate for the development of small to moderate size models The latest version can be retrieved from the archive Monitoring Displays The Soar group at USC ISI have created several displays of working memory elements using raw graphic system calls These displays have not been generally made available but are likely to end up in the SDE or the general Soar release These include monitor windows that show the goal stack as the model runs These indicate that those who are sufficiently familiar with Soar C and X windows can create their own displays and that such displays are useful For a brief period in 1992 the DSI Ritter amp Larkin 1994 in press was the most complete and used interface for Soar if you counted use of its components It is now unused because it ra
25. mmunicate with numerous extensions the Tcl archive contains over 250 extensions These extensions include graphing utilities a socket interface and a hypertext system e Portable to Unix Mac running a Unix OS and PC using existing Tcl interpreters e More elaborate and context dependent aliases Interpreted language with libraries of extension modules TclSoar Schwamb 1994 described in Table 3 is a version of Soar incorporating a Tcl an embeddable extendible command line interpreter TclSoar also includes the Tk extension of widgets graphical objects that enables users to create graphical extensions interactively To find out more about TclSoar Tcl and Tk see the web documentation http www isi edu soar schwamb TclSoar is available from the archive In addition to providing a more extendible command line interface TclSoar makes it easier to connect graphical interactors and simulations to Soar It provides access to a simple simulation language existing Tcl simulations and an additional model simulation interprocess control approach While it will be a master at none of these its greatest potential is serving as a uniform and integrating environment for incorporating additional displays of the running model creating simple simulations easily and hooking up more complicated simulations routinely Using existing simulations We outline several existing simulations to further define our needs and review possibilities th
26. mulation may not see time pass equally Multi agent Soar Hucka personal communication 1994 was created was to get round this non deterministic scheduling for multiple models by running them synchronously i e so that they each see each model cycle at the same time Unix Currently machines that routinely run Unix as their operating system are generally the fastest and most commonly available for developing cognitive models and simulations Their only drawback is that they cannot automatically provide reliable timing information about user s input As a multitasking operating system with multiple processes and a scheduler a user action is not necessarily timestamped when it is input Delays of 10 ms to several seconds are typical For example although time stamps with 200 ms accuracy are quite likely with an available GNU Emacs package log el by Erik Altmann incorporated into Dismal a significant portion of the keypresses perhaps 5 take 0 ms It is not possible to have a single user process running under the Unix operating system in a routine or simple manner However timing glitches are minimised if the simulation process runs at maximum priority timestamps incoming keystrokes first enough swap space is provided and some other minor resource issues are resolved Another option is having a small high priority process that just received keystroke data timestamped it and passed to the simulation There also are real time Unix kernels
27. n only in the Lisp based version of Soar Soar5 but it does offers several lessons for modelling Its graphic displays of the internal state of Soar showed how useful such displays can be for debugging and teaching e g it was used at demos and to help create the Soar video Lehman Newell Newell Altmann Ritter amp McGinnis 1994 Its functionality is slowly being taken up by the SDE and the monitoring windows Its only major drawback was that it slowed down Soar by approximately a factor of two This was caused by two things a It was implemented in an older larger version of Garnet that made the image large enough to swap and b it included a continuous graphic trace Sadly any graphic display running on the same processor as a model will slow it down Table 3 Some of the features of TclSoar e Full command language including variables data structures command structures and file operations e Tk extension enables interactive Graphical User Interface GUI creation and a simple process communication mechanism called send GUIs can be created and modified without C coding or recompilation Data from Soar can be sent to other applications running on other machines if desired such as a graphing utility to generate run time performance displays e Script execution from productions Hence a model can call any user command This enables users to run Tcl Tk scripts when a particular Soar event occurs e Ability to add or co
28. r PC There are other platform independent toolkits for creating graphical interfaces so we include a sample of the most likely candidates What we do not cover are environments that our modelling language does not have specific affinity for that we don t have access to or that we cannot add easily to our environment Table 4 Possible advantages to using an existing simulation e Savings in development time e Community of existing users Possible existing analyses and summaries of behaviour Possible existing models of behaviour for comparison Garnet Garnet Myers et al 1990 calls itself a graphical user interface management system but it is also an object oriented graphical interface toolkit and environment It is currently written in Common Lisp and runs on Unix and the Macintosh It is actively being developed at CMU by the Garnet group While bug responses are good there is lots of help because it comes with source code and there is a Usenet bulletin board comp windows garnet for users Garnet is available via anonymous FTP from a gp cs cmu edu 128 2 242 7 You must directly change to the distribution directory with cd usr garnet garnet Get the README file there and follow its directions A C based Garnet is under development but it is not clear the implications of this as Garnet takes advantage of many Lisp features including its interpreter This version will be developed to be portable across Unix M
29. to make errors Providing an additional type of data to understand behaviour and its evolution e How do they use and scan displays e How tightly can a model be fitted to human data in these areas learning reaction times and sequential behaviour Table 2 notes several necessary features of such a task to study the evolution of rapid decision making Our approach with this paper We outline here the existing simulation software that is most likely to meet these requirements We start with our preference for a particular environment for developing Soar models perhaps the most difficult aspect of this task and point out the useful features that should be provided by model development environments This influences the choice of machine and thus the available simulations and the communication medium between the model and the simulation Figure 2a The tower of Nottingham Figure 2b The pieces that make up each level in the Tower of Nottingham Figure courtesy of H Wood Model development environments The environment for developing the model is an important decision perhaps the most important The analyst needs to be able to see the model understand how it works and modify it We base our discussion around cognitive models in Soar While some may be directly interested in Soar this discussion should be seen primarily as a discussion of the types of interfaces that can be provided for cognitiv
30. ubjects can interact with Raw Xlib Window calls in C It will often be possible to implement a simulation in the language used to implement the model Simulations created in the foundation language are easy to connect to the model For Soar it is C It is possible to use C and raw calls to the Xlib window functions to create a simulation and graphic display under the X window system Jones 1989 Schwamb at ISI personal communication 1994 recommends this approach where speed is an issue for it produces the fastest simulation because it is linked into the native display system It will perhaps be the slowest way to develop because little support is provided and it requires the most skill to implement Other simulation languages A wide listing of platform independent graphical user interface building tools is available as a frequently asked questions FAQ posting to comp windows misc also news answers by Guthrie wade nb rockwell com This FAQ and ones on Lisp and window systems often posted to comp lang lisp were used to help derive the following other candidates Visual Basic Visual Basic Anderson 1994 is a user interface rapid prototyping language that comes with widgets that can have structured BASIC code attached It is possible to create good interfaces fairly quickly The downside is that the code is hard to manage Visual Basic is not free but not expensive a few hundred pounds SUIT Simple User Interface Toolkit SUIT
31. x 8202 North Carolina State University Raleigh NC 27695 tel 919 515 7447 fax 919 515 2682 for 175 or 250 for both DOS and Mac There is a users group ctug athena mit edu GNU Emacs Lisp GNU Emacs has a rather open architecture including a Lisp extension language It would be straightforward to create a simulation in GNU Emacs Lisp and tie it to a model using the process control in SDE Hucka 1994 This would not provide a fully graphical display as we need for our tasks and the size of the simulation would be limited by the speed of this Lisp but it would allow a model to talk to a piece of software with a textual interface such as the Dismal spreadsheet Nichols amp Ritter in press Model based _ simulations It is often possible to create a simulation in the modelling language itself In Soar these range from productions directly implementing the task to more sophisticated systems for accessing external timers and files Guttman amp Huffman 1992 Nelson 1994 These approaches are useful for fixed sets of inputs and simple interactions such as the Tower of Hanoi but they are not powerful enough to implement either of our tasks This approach can be awkward because modelling languages are not designed to be simulation languages The approaches one must take are often not supported by debuggers and visual displays because they use the language in odd ways and it is typically not possible to create a display that s
Download Pdf Manuals
Related Search