Agent-based Simulation Platforms: An Updated
Contents
1. has changed substantially The Repast Simphony version North et al 2007 http repast sourceforge net the current version of which was released in 2010 uses a new conceptual approach and is in important ways a different platform from previous versions Part of CC Creative Commons License 2011 2 Author First Name Last Name the Simphony version is ReLogo described by its developers see http repast sourceforge net repast_simphony html as a dialect of Logo Logo being a computer language designed in the 1960s primarily for education Papert 1980 ReLogo is clearly based on NetLogo as it includes almost of NetLogo s primitives and many of its graphical interface tools NetLogo and ReLogo also share a common goal of enabling novice programmers to develop agent based models According to NetLogo s developers NetLogo follows Logo s philosophy of ease of use providing a low threshold of entry for new users Tisue and Wilensky 2004 Similarly the developers of Repast Simphony strongly encourage users with limited programming background to begin model development with ReLogo http repast sourceforge net docs html In addition to the ease of use of ReLogo as compared with the full blown version of Repast Simphony some potential benefits of ReLogo appear to be 1 allowing Repast users to take advantage of the many powerful NetLogo primitives and programming concepts such as built in grid cell an2. a model so display updates are not too fast to observe The experiments were conducted with standard installations of each platform on a Windows 7 computer We also compared execution speed with display updates turned off In NetLogo this was accomplished using the no display primitive and the BehaviorSpace tool in ReLogo the Parameter Sweep tool was used which automatically turns off the model display 3 Results 3 1 NetLogo evaluation Version 5 of NetLogo has few fundamental differences from Version 2 and in fact all the StupidModel code we wrote for Version 2 ran in Version 5 without modifications other than those made by NetLogo s automatic translator In re implementing StupidModel we noted a number of changes in NetLogo relevant to criticisms we made in 2006 e Default scheduling of an agent action is now for agents to execute one at a time instead of with pseudo concurrency This change makes it possible to know the exact order in which events occur and avoids the potential for rather spectacular artifacts of concurrency e g Section 14 2 of Railsback and Grimm 2012 Now NetLogo by default randomly shuffles any list of agents before they execute an action but there is also a well documented way to execute agents in order of any of their variables e The graphical interface now provides Inputs a widget that allows the exact value of a global variable or parameter to be entered e There is now a built in
3. blurred when experiments are conducted in NetLogo using the BehaviorSpace tool In BehaviorSpace simulation experiments operation of the widgets on the user interface is taken over by BehaviorSpace A BehaviorSpace experiment is created by filling out a dialog to define parameter sweep values outputs to be reported and code to be performed when a model run is initialized and finished In ReLogo a model is broken up into at least 2 parts Groovy code and XML files The breakdown between these components of the model is not as intuitive as in NetLogo Groovy code is written to perform the execution tasks and to define some of the variables XML files are used to define other parameters in a model such as the size of the model s space But other characteristics of the space such as whether or not it should be toroidal are again defined in Groovy code Similarly the definition of User Interface widgets is also done programmatically We find this division between Groovy code and XML to be suboptimal 3 3 Execution speed comparison ReLogo consistently took 20 times longer or more to run our models than did NetLogo when the model display was turned on Table 1 Turning off the model display caused both platforms to run somewhat faster with NetLogo running 3 20 faster and ReLogo running 3 30 faster As would be expected the largest gains in speed are for the simplest models versions 1 and 3 in Table 1 and the smallest gains are for the most
4. EMCSR pp pp year ISSN XXXX XXXX http conference emcsr net Agent based Simulation Platforms An Updated Review Steven L Lytinen School of Computing College of Computing and Digital Media DePaul University 243 S Wabash Room 645 Chicago IL 60604 USA lytinen cs depaul edu 01 312 362 6106 Steven F Railsback Lang Railsback amp Associates 250 California Avenue Arcata CA 95521 USA Steve LangRailsback com Abstract We review and evaluate two related agent based simulation platforms version 5 0 of NetLogo and the ReLogo component of Repast We implemented the StupidModel series of 16 pseudo models in both platforms these codes contain many elements of basic agent based models and can serve as templates for programming real models Subsequent to the similar review we published in 2006 NetLogo has evolved into a powerful platform for scientific modeling while retaining its basic conceptual design ease of use and excellent documentation ReLogo is a new component of Repast it implements NetLogo s basic design and its primitives in the Groovy programming language embedded in the Eclipse development environment ReLogo successfully reimplements much of NetLogo and its translator was successful in converting NetLogo codes into ReLogo Overall we found ReLogo considerably more challenging to use and a less productive development environment Using ReLogo requires learning Groovy and Eclipse and becoming familiar with
5. J 2005 Eclipse Rich Client Platform Designing Coding and Packaging Java Applications Boston MA Peason Education North M Howe T Collier N and Vos R 2007 A declarative model assembly infrastructure for verification and validation In S Takahashi D L Sallach and J Rouchier Eds Advancing Social Simulation The First World Congress Heidelberg Springer Papert S 1980 Mindstorms Children Computers and Powerful Ideas New York Basic Books Railsback S F Lytinen S L amp Jackson S K 2006 Agent based simulation platforms review and development recommendations Simulation 82 9 609 623 Railsback S F and Grimm V 2012 Agent based and individual based modeling A practical introduction Princeton Princeton University Press Sondahl F Tisue S and Wilensky U 2006 Breeding faster turtles Progress toward a NetLogo compiler In Proceedings of the Agent 2006 Conference on Social Agents Results and Prospects pp 5 16 Chicago USA CC Creative Commons License 2011 10 Author First Name Last Name Standish R 2008 Going stupid with EcoLab Simulation 84 12 611 618 Tisue S and Wilensky U 2004 NetLogo A simple environment for modelling complexity Presented at The International Conference on Complex Systems Boston Massachusetts May 2004 Wilensky U 1999 NetLogo http ccl northwestern edu netlogo Center for Connected Learning and Computer Based Mod
6. Repast s complex organization documentation and learning materials are far less abundant and mature than NetLogo s Though we did not investigate thoroughly it is not clear what kinds of models could readily be implemented in ReLogo but not NetLogo On average NetLogo executed our example models approximately 20 times faster than ReLogo Keywords agent based simulation platforms NetLogo Repast ReLogo template models Acknowledgement We thank Jonathan Ozik Argonne National Laboratory for assisting us with our model implementations in ReLogo 1 Introduction Agent based models ABMs continue to increase in importance and popularity as a way to study complex systems so the software platforms designed for them are also increasingly important In 2006 we published a review of five platforms looking at their conceptual basis the experience of programming a series of example models and informally execution speed Railsback et al 2006 That review identified NetLogo Wilensky 1999 http ccl northwestern edu netlogo as being especially well designed and documented and easiest to learn and use while Repast was one of the platforms that we recommended for models that are especially demanding computationally or not well fitted to NetLogo s conceptual style Several new versions of NetLogo have made important changes that address many of the limitations we identified in 2006 though its overall design has been stable Repast however
7. ained stable NetLogo is now a powerful tool widely used in science and we recommend it strongly especially those new to modeling and programming but also for serious scientists with software experience NetLogo was chosen by Railsback and Grimm 2012 as the platform for their textbook on scientific agent based modeling CC Creative Commons License 2011 8 Author First Name Last Name 4 2 ReLogo ReLogo is the latest in a series of directions taken by the Repast project Since our 2006 review the original Java version and a Python language version with drag and drop model building appear to have been abandoned in favor of the new Java and Groovy based Simphony version that includes ReLogo The main critiques of Repast we made in 2006 high complexity not mitigated by either strong organization or good documentation and lack of a conceptual framework still seem valid A NetLogo user would notice a number of differences upon trying to use ReLogo e In addition to understanding the NetLogo primitives a working knowledge of Groovy and possibly Java is necessary e Working in Eclipse is far more complex and less efficient in many ways than the NetLogo environment e The necessity to understand and keep track of low level programming issues such as Eclipse projects and Java Groovy packages classpaths dependencies import statements etc is new and challenging and e ReLogo requires following object oriented p
8. cember 8 2011 This specification is available at http condor depaul edu slytinen abm StupidModelFormulation 2011 pdf CC Creative Commons License 2011 EMCSR pp pp year 3 such changes Hence StupidModel requires a few programming statements that appear clumsy compared to others The NetLogo implementations were completely independent of those conducted for our 2006 review of NetLogo version 2 and attempted to make as much use as possible of NetLogo s style and built in primitives Because ReLogo is based on NetLogo the ReLogo implementations of StupidModel were based on the NetLogo code the same code design and primitives were used as much as possible We evaluated NetLogo primarily by how it has changed since the version 2 we evaluated in 2006 ReLogo was evaluated by noting parts of the programs that seemed particularly easy or difficult to program and important ways that it differs from NetLogo 2 2 Execution speed Execution speed was evaluated for seven versions of StupidModel The model was run for 1000 time steps with the code using NetLogo s and ReLogo s timer primitives to report elapsed time between the first and last time steps The time to initialize models by executing their setup procedure was not included These experiments were conducted with the graphical displays active Both platforms were set to run as fast as they could while updating the display once per time step NetLogo allows users to slow
9. ches in later versions patches grow food which bugs eat bugs optimize their movement according to food availability bugs reproduce and die and in the last version predators are introduced StupidModel has since 2006 also been implemented by others in at least four other platforms including Repast Simphony http code google com p repast demos wiki StupidModel EcoLab Standish 2008 Behavior Composer and Xholon see www swarm org index php Software templates In our current study we updated the StupidModel specifications to clarify a number of small ambiguities Then we implemented all 16 versions in both NetLogo and ReLogo All code is available at http condor depaul edu slytinen abm During re implementation we realized that several parts of StupidModel would be much more natural to implement in NetLogo and ReLogo and probably faster to execute if changed in small ways The most prominent example would be to have agents interact with grid cells within a circular instead of square neighborhood However to maintain comparability with other implementations and reviews based on them we did not make 1 As of 2006 the NetLogo developers were working on a compiler to translate NetLogo code directly into byte code by passing the additional interpretation phase see Sondahl Tisue and Wilensky 2006 for details In NetLogo v 5 some primitives are compiled while others must still be interpreted Seth Tisue e mail message to author De
10. complex model version 15 since proportionally more time is devoted to updating the display in simple models The somewhat unexpected result that NetLogo is much faster than ReLogo is probably due at least in part to NetLogo implementing its primitives more efficiently than they are in ReLogo We did not attempt to compare the platforms implementation of their primitives by reading source code However we did conduct some simple experiments replacing complicated primitive statements with simpler low level code such changes tended to make ReLogo faster and NetLogo slower which is consistent with the assumption that NetLogo primitives are implemented efficiently Because NetLogo is its own language its compiler may be able to perform optimizations that are impossible in ReLogo Another contribution to NetLogo s speed is the aforementioned partial compilation of NetLogo code directly into Java byte code 4 Conclusions and Recommendations 4 1 NetLogo Through a steady series of revisions the current version 5 of NetLogo has addressed all of the critiques and wish list items in our review of version 2 Railsback et al 2006 with the exception of adding a stepwise debugger Many of these changes Sect 3 1 substantially improved NetLogo s usefulness as a scientific tool a key example is ability to set up and execute simulation experiments with ease and speed using BehaviorSpace Yet NetLogo s core concepts and overall design have rem
11. d agent classes and agentsets 2 providing a link between NetLogo and Repast so users familiar with NetLogo can transfer their models to Repast and then use other features of Repast not available in NetLogo and 3 potentially better execution speed because ReLogo models are written in the language Groovy which compiles directly into Java byte code while NetLogo uses its own language that must be interpreted This study updates Railsback et al 2006 using similar methods to review the newest versions of NetLogo version 5 0 and ReLogo part of the 2 0 beta version of Repast Simphony released December 2010 From the review we draw conclusions on the suitability of NetLogo for scientific ABMs and on how well ReLogo provides the three potential benefits identified above 2 Methods To make this study as comparable as possible we use methods very similar to those of our 2006 review Programming experience was evaluated by implementing a series of ABMs in both platforms and we informally evaluated execution speed for several of those models 2 1 Programming experience in implementing StupidModel In Railsback et al 2006 we evaluated ABM platforms by programming 16 versions of StupidModel a pseudo model designed to test common software tasks of agent based modeling and also to provide a template from which real ABMs can be adapted The simplest version of StupidModel consists of bugs moving randomly about a 100x100 world of pat
12. eling Northwestern University Evanston Illinois About the Authors Steven Lytinen Steve Lytinen PhD Yale University is a professor in the School of Computing College of Computing and Digital Media at DePaul University His research expertise is in artificial intelligence and agent based simulation Steven Railsback Steve Railsback PhD University of Bergen Norway is an adjunct professor in the environmental modeling program Department of Mathematics Humboldt State University and an environmental engineer with Lang Railsback amp Associates His expertise is in individual based ecological modeling and river management CC Creative Commons License 2011
13. ers to select how many of their computer s processors to use Individual model runs are sent to various processors but results are all gathered in one output file Hence simulation experiments with multiple model runs can be made parallel with almost no effort e NetLogo is now open source so users can look at and modify its source code at https github com NetLogo Its documentation remains extraordinarily complete and accurate compared to that of other platforms so there is also much less need to read source code Only one potential improvement suggested in our 2006 review a stepwise debugger has not been added to NetLogo 3 2 ReLogo evaluation 3 2 1 Programming experience A fundamental design difference between NetLogo and ReLogo is that NetLogo provides a programming language designed specifically for agent based modeling while ReLogo models are written in Groovy Koenig et al 2007 htip groovy codehaus org a lightly typed dynamic object oriented general purpose programming language We found it unlikely that ReLogo could be used productively without at least a rudimentary knowledge of Groovy Thus ReLogo falls into the category of framework and library family of modeling platforms Railsback et a 2006 There is a trade off between these two approaches For novice programmers the NetLogo language appears easier to learn since its syntax is simpler than the syntax of an object oriented language like Gro
14. etLogo being relatively stable and gradually refined over a number of years giving its developers the chance to implement its primitives and graphic displays efficiently Another potential explanation for the speed difference is that the Windows version of NetLogo is packaged with a server version of the Java virtual machine which faster than the default virtual machine which Eclipse is typically configured to use However our experiments with NetLogo using both virtual machines server and default indicate that the server versions could explain only a 10 20 increase in execution speed 4 3 Potential benefits of ReLogo to Repast users In Sect 1 we speculated that two motivations for the development of ReLogo were to allow Repast users to take advantage of NetLogo s primitives and concepts and to facilitate the transfer of models from NetLogo to Repast so unique features of Repast can then be used Are these benefits provided by ReLogo We found ReLogo to be successful in the sense that it does implement NetLogo s fundamental concepts a world of grid patches and turtle agents and primitives This would be beneficial compared to just using NetLogo if ReLogo and Repast could then be used to do things that seem impossible or very clumsy in NetLlogo Common examples of such things are multiple spaces and non square spatial units the trout model of Railsback et al 2009 represents multiple sections of river as separate spaces each made
15. getBugSize return 1 else return 15 turtles for Turtle t sTurtles t move The sortBy primitive in ReLogo takes 2 parameters a 2 argument in this case a and b closure that returns an integer and an AgentSet of turtles The returned integer should be 1 0 or 1 depending on whether a gt b a b or a lt b 3 2 2 Documentation NetLogo is a much more mature modeling platform than ReLogo and its documentation is far superior to ReLogo s An extensive User Manual and many example codes are provided with NetLogo In the NetLogo programming environment a right click on any primitive brings up the entry for that primitive in the NetLogo dictionary Most primitive dictionary entries describe the syntax of the use of the primitive a description of what the primitive does and an example of how the primitive is used For example here is the NetLogo dictionary entry slightly modified for the max one of primitive max one of agentset reporter Reports the agent in the agentset that has the highest value for the given reporter If there is a tie this command reports one random agent with the highest value Example max one of patches count turtles here A reporter in NetLogo is a procedure which returns an answer i e a function The NetLogo dictionary entry clearly shows that the reporter must be enclosed in square brackets On the other hand the ReLogo dictionary tends to be less informative Here is t
16. he corresponding ReLogo dictionary entry for the same primitive maxOneOf List a Closure reporter CC Creative Commons License 2011 6 Author First Name Last Name Returns the turtle with the largest value when operated on by a set of commands The dictionary entry does not describe the syntax of closures in Groovy and does not give an example of how to use maxOneOf 3 2 3 Development environment The NetLogo development environment is very intuitive for a novice user Figure 1 NetLogo has a syntax checker button the button with a check mark icon that the user can click to ensure that code in the model is syntactically correct This can almost instantly identify most errors in code without running the model StupidModel1 0 NL5 NetLogo C Pubs Lytinen ABMs StupidModel NetLogo5 of x StupidModel10 NL5 NetLogo C Pubs Lytinen ABMs StupidModel NetLogo5 ojx File Edit Tools Zoom Tabs Help File Edit Tools Zoom Tabs Help InterFace Info Code Interface Info Code f n I view updates 2 S 8 ic BOTS F l Settings aa en Procedures l I Indent automatically te Add normal speed onticks A pa e Model o update food patch procedure Mo ti dre _ ticks 58 set food available food available random float max food prod end to move let empty neighbor one of stupid neighbors if empty neighbor nobody C move to empty neighbor to grow a turtle primitive Model 2 cons
17. ments For example let s say we want to find the neighbors of one of the turtles named Ralph Here is the NetLogo statement to do this let ralph neighbors neighbors of one of turtles with name Ralph CC Creative Commons License 2011 EMCSR pp pp year 5 In ReLogo this is written def ralphNeighbors neighbors of oneOf turtles with name equals Ralph Among other uses curly braces in Groovy are used to define closures a concept familiar to computer scientists but most likely not to the novice programmers that ReLogo is designed for A closure is an unnamed procedure which is passed one or more parameters and returns an answer In this case the closure name Ralph is passed each of the agents in list turtles and returns true or false depending on whether or not a turtle is named Ralph Other primitives which by their nature require more complexity are considerably different in NetLogo and ReLogo For example here is a snippet of code from one of our StupidModels in which agents execute their procedure move in order sorted by their size so that bigger agents move first In NetLogo the code is foreach sort by bug size of 1 gt bug size of 2 turtles ask move Because sorting involves comparing pairs of items NetLogo uses 1 and 2 to refer to the pair In ReLogo we had to write this as follows def sTurtles sortBy a b gt if a getBugSize lt b
18. ovy On the other hand because ReLogo models are written in a general purpose programming language it could potentially be easier to write models that require complex actions by agents which are not captured by the primitives of the modeling platform One additional potential benefit of ReLogo is the ability to also use Repast Simphony s larger library and wider variety of modeling capabilities However it is not immediately clear how this would be done since Repast Simphony models do not use turtles and patches Two fundamental concepts of Logo languages are built in spatial grid cells known as patches and mobile agents known as turtles NetLogo and ReLogo both share the notion of primitives i e agent patch actions or other programming constructs which are built in to the language For some primitives the difference in syntax between the platforms is negligible For example a commonly used primitive in both languages is neighbors which reports the neighboring patches to a turtle or a patch An example of this primitive s use in NetLogo is the following statement which creates a new local variable containing an agent s neighbors let npatches neighbors And in ReLogo def nPatches neighbors A key strength of NetLogo is the ability to combine simple primitives such as neighbors into compound statements that are very powerful and still easy to read ReLogo significantly complicates the syntax of such state
19. rogramming conventions such as having code for each class turtles patches the observer etc in separate files and not having global variables Table 1 Execution time comparison Number of agents and runtimes represent the mean of 5 runs executed in each platform Description of StupidModel version Number NetLogo ReLogo Speed ratio of agents runtime runtime NetLogo secs secs ReLogo speed Version 1 The world has 100 x 100 100 6 3 173 27 7 patches At each tick each bug moves to a randomly chosen vacant patch with x and y co ordinates 4 from current location V 3 Patches grow food After moving 100 9 1 189 20 9 bugs eat food from their patch up to a maximum rate of consumption Turtles grow in size based on consumption V 11 Bugs select patch with maximum 100 9 7 191 19 7 food availability Larger bugs move first V 12 Bugs produce up to 5 offspring 140 13 5 280 20 7 and die when they reach a certain size V 15 The world size is 256 x 112 Patch 1330 103 2509 24 3 food growth rates are read from a file V 16 200 predators now eat bugs At 729 65 1 1516 23 3 each tick predators move to an immediately neighboring patch that contains a bug provided no other predators are in the patch then eat the bug CC Creative Commons License 2011 EMCSR pp pp year 9 The difference in execution speed between ReLogo and NetLogo could largely be a result of N
20. tain a method that is called in a ReLogo model the mistake is not identified by the Groovy compiler Similarly incorrect use of a variable name i e it is not associated with the appropriate class in some circumstances is not be caught by Eclipse or the compiler and will only be flagged as an error at runtime Our experience with ReLogo is that it takes half to a full minute to start running a model to expose the errors Hence having errors caught only at runtime negates much of the benefit of Eclipse and significantly slows code development CC Creative Commons License 2011 EMCSR pp pp year 7 3 2 4 Code vs XML vs interface dialogs In NetLogo there is generally a clear cut intuitive division between portions of a model written in code vs those portions developed via menus and dialogs on the interface Generally speaking procedures and variables are written as code under the Code tab in the NetLogo user interface GUI Widgets such as sliders monitors inputs graphs and histograms are defined under the Interface tab using menus and filling in dialogs While this makes it easy to develop a user interface for a model this approach has the disadvantage that there is no single place where one can look to see everything that the model contains In fact version 5 provides the less clear cut option of writing code to update plots inside the plot widgets instead of on the code tab This division of labor becomes somewhat
21. umption min list 1 0 food available Model 3 sumption min list bug max cons food available Model 5 umption food consumption Food may bec color goes from black t StupidModel goes whit red set color 10 scale color red bug size 10 0 2 end Command Center to report stupid neighbors a turtle or patch reporter ERAEN ee Figure 1 NetLogo s graphical interface and code tabs In contrast ReLogo is packaged within the Eclipse integrated development environment http www eclipse org Eclipse is sophisticated and very powerful but for a novice there is a steep learning curve for learning Eclipse as is suggested by the large number of textbooks devoted to the Eclipse IDE see for example McAffer and Lemieux 2005 as well as Groovy and the ReLogo primitives While we recommended Eclipse as an environment for developing with framework and library platforms Railsback et al 2006 the complexity of ReLogo in Eclipse stands in stark contrast to NetLogo Another usability disadvantage of ReLogo results from Groovy being a loosely typed language The Groovy compiler in Eclipse at least as downloaded from the Repast Simphony Web site cannot identify many of the kinds of programming errors that are identified in NetLogo by its syntax checker or identified by Eclipse in Java code In ReLogo many syntax errors are only identified at runtime For example if a class does not con
22. up of rectangular or in later versions polygonal cells While such models could almost certainly be implemented in Repast it is not clear that ReLogo would be very useful in doing so Many of the NetLogo primitives and concepts are specialized for a single square grid world and do not make sense in other contexts Hence ReLogo may have limited usefulness for models that could not easily be implemented in NetLogo Another successful aspect of ReLogo is its tool for importing NetLogo models and translating them into ReLogo code and XML files In fact we found given the state of ReLogo documentation one of the most effective ways to learn how to program ideas in ReLogo was to first program them in NetLogo and then have the translator convert the code to ReLogo After developing some familiarity with Groovy we found ReLogo easier for implementing StupidModel than Repast We implemented some versions of StupidModel in Repast Simphony However the question remains whether any real benefit over using NetLogo is likely to result from translating a model to ReLogo Before doing so we would recommend making sure that there are features of Repast that are not available in NetLogo or in NetLogo s many extensions and will work with models using ReLogo s structure and primitives References Koenig D Glover A King P Laforge G and Skeet J 2007 Groovy in Action Greenwich Connecticut Manning Publications McAffer J and Lemieux
23. variable for the number of time steps executed and primitives for re setting and reporting it e The documentation now clearly says when newly created agents execute their initial behaviors Code to create a number of new agents first creates all of them then has each new agent execute its initial behaviors e The dimensions of the space can now be set from the code That allowed us for versions 15 and 16 of StupidModel to read an input file of patch data determine the dimensions of the space from the file set the space to those dimensions and then re read the file to set its patch variables This capability does not appear to be present in ReLogo CC Creative Commons License 2011 4 Author First Name Last Name e The space can be very easily switched from toroidal to non toroidal and primitives that depend on the difference automatically behave appropriately for the current setting e The graphical interface not only includes a checkbox to turn off graphics updates but also provides a speed controller that lets the user decide how often the display is updated and therefore how much computation resource is available to execute the model instead e It is possible to keep code in multiple files and on multiple pages in the code editor though code still need not be separated by class e NetLogo s BehaviorSpace tool which can be used to run replicates and parameter sweeps of models was significantly improved by allowing us
Download Pdf Manuals
Related Search