A Stigmergy Collaboration Approach in the Open Source Software ...

A Stigmergy Collaboration Approach in the Open Source SoftwareDeveloper CommunityXiaohui Cui, Laura L. Pullum, Jim Treadwell and Thomas E. PotokAbstract— The communication model of some self-organizedonline communities is significantly different from thetraditional social network based community. It is problematicto use social network analysis to analyze the collaborationstructure and emergent behaviors in these communitiesbecause these communities lack peer-to-peer connections.Stigmergy theory provides an explanation of the collaborationmodel of these communities. In this research, we present astigmergy approach for building an agent-based simulation tosimulate the collaboration model in the open source software(OSS) developer community. We used a group of actors whocollaborate on OSS projects through forums as our frame ofreference and investigated how the choices actors make incontributing their work on the projects determines the globalstatus of the whole OSS project. In our simulation, the forumposts serve as the digital pheromone and the modified Pierre-Paul Grasse pheromone model is used for computing thedeveloper agents’ behavior selection probability.HI. INTRODUCTIONierarchical structure and top-down decision-makinghave dominated human society’s structure andfunction. In a hierarchically structured society, collaborationbehaviors are governed by policies and laws, and people areexpected to do what they are assigned and wait forinstructions. However, some communities are known todisplay a large amount of distributed and bottom-upstructure. In these communities, the social patterns thatpeople form are often organized without explicit leaders,chains of command, or fixed communication networks.Examples of such spontaneously emerging social groupsinclude fans at a sport stadium, grassroots politicalmovements, terrorist networks, Wikipedia authors, and opensource software (OSS) developer communities. Tounderstand how the collaboration or work gets done in thesegroups, a new approach called social network analysis orSNA, has been used as an effective method for revealing theconnections and patterns. SNA is based on an assumption ofthe importance of relationships among interacting units. ItManuscript received May 1, 2009. This work was partially funded byLockheed ShareVision research funds, Oak Ridge National Laboratory SeedMoney funds and internal Lockheed research funds.Xiaohui Cui is with the Computational Sciences and EngineeringDivision, Oak Ridge National Laboratory, Oak Ridge, TN 37831, USA(phone: 865-576-9654; fax: 865-214-0337; e-mail: cuix@ornl.gov).Laura L. Pullum is with Lockheed Martin Corporation, St. Paul, MN55164-0525, USA (e-mail: laura.l.pullum@lmco.com).Jim Treadwell and Thomas E. Potok is with the Computational Sciencesand Engineering Division, Oak Ridge National Laboratory, Oak Ridge, TN37831, USA.examines the structure of social relationships in a group touncover the informal connections between people andprovide both a visual and a mathematical analysis ofcomplex human systems.However, the communication model of some onlinecollaborative groups is significantly different from thetraditional social network-based collaboration community.In the traditional community, a message (e.g., a spokencommunication) is directed to a particular individual orindividuals at a particular time. But such peer-to-peercommunications are very rare in those online collaborationgroups. For instance, Wikipedia is an online collaborativeencyclopedia created and edited by volunteers from aroundthe world. Anyone can add or change an entry in Wikipedia.However, there are no direct communications between thosecontributors. Audris et al. [1] point out that OSS developersrarely if ever meet face-to-face or even via telephone.Posting messages on forums or via email lists are the majorinformation exchange methods for OSS developers andmany other online collaborative groups. These messagesserve as a shared medium for storing information so that itcan be interpreted by other individuals later. Most of thetime, the message sender does not know by whom and atwhat time these messages will be viewed and the messageconsumers typically do not care who sent the message.Without clearly defined informal connections and messageflows between individuals, SNA will have difficultyunderstanding the emergent behaviors represented in theseonline communities.In this research, we propose a stigmergy approach foranalyzing the collaboration architecture in these onlinegroups. Given the significant amount of detailed informationavailable on OSS communities’ internal structure, OSSdevelopment communities offer a unique opportunity tounderstand the large self-organized online entity. In thispaper, we present a stigmergy agent based model forexamining how the behavior of individuals affects theemergence of an OSS project’s global status.II. STIGMERGY COLLABORATION IN OSS COMMUNITYThe term stigmergy was first proposed by Pierre-PaulGrasse in the 1950s in his research on termites [2]. Grasseshowed that a particular configuration of a termite colony’senvironment triggered a termite to modify its environment(i.e., drop mud in a particular place for building ormaintaining the nest). The modification in turn stimulates

the original or other termites in the colony to furthertransform their environment. Grasse defined stigmergy as:“the stimulation of the workers by the very performancesthey have achieved” [2]. The stigmergy process has beenobserved in termites, ants, bees, and wasps in a wide rangeof activities. In a termite colony, a highly complex termitenest is not caused by the net building knowledge ofindividual termites. It is the result of the collective behaviorof large numbers of individual termites performingextraordinarily simple actions in response to their localenvironment. There are no direct communications betweentermite workers for coordinating their nest building actions.The modified environment caused by an individual termite’ssimple actions serves as the coordinating signals. The stateof the nest structure triggers some behaviors, which thenmodify the nest structure and trigger new behaviors until theconstruction is complete.The OSS developer community is a new kind of onlinesoftware development group where participants can read,modify, and redistribute software source code without cost.In the software development field, a long-held belief is thatlarge scale software development should be regarded as acollaborative activity to be conducted in a hierarchicallystructured organization such as a company. However, manyOSS development systems do not evolve the same as theseproprietary systems. The distributed and unplanned OSSdevelopment model has been shown to be very effective as asoftware development paradigm and outperforms theproprietary software development schemes. In an OSSdeveloper group, while people are much more intelligentthan social insects, open software development usesessentially the same stigmergic mechanism for collaborating.Participants in OSS projects mainly engage in onlinediscussion forums or threaded email messages as a centralway to observe, participate in, and contribute to publicdiscussions of topics of interest to ongoing projectparticipants [3]. The newly developed software source codesare uploaded to the community website for the purpose ofbeing scrutinized by the members of the community. Anybug, error or lacking functionality will be pointed out andthus entice community members to take up the shortcoming.The concept of stigmergy provides a theory for explaininghow disparate, distributed, ad hoc contributions fromindividuals could lead to the emergence of the largestcollaborative enterprises the world has seen.III. RELATED WORKIn most online hosting environments, project relatedactions are logged and the log information can later bemined to understand the community structure and interactionpatterns. This log data provides enormously detailedinformation for analysis [4, 5]. Crowston, et al. [6] proposeda model for effective work practices in OSS development.The model was based largely on an existing model of groupeffectiveness initially proposed by Hackman [7] in 1986.Smith, et al. [8] presented an agent-based OSS simulationmodel that includes the software modules’ complexity, thesoftware’s fitness for purpose, the motivation of developers,and the role of users in designing requirements. In researchon how OSS developers collaborate, research considers theOSS movement as a self-organizing system and acollaborative social network [9]. SNA was used foranalyzing OSS community structure. Actor relationships arerepresented as nodes and links. The actor can be a user ordeveloper. Every node i represents an actor within thenetwork; link(i,j) denotes a social tie between actors i and j.However, as we mentioned in the introduction, in the OSScommunity, direct connections between actors as unusual.They exchange information through the forum or email-listindirectly. Most of time, the actor does not even know whothe recipient of his/her message is prior to sending themessage. Hence, SNA may not be able to explain the forumbased collaboration concept characterizing the OSScommunity.Elliott [10] argued that collaboration in small groups(roughly 2-25) relies upon social negotiation to evolve andguide its process and creative output. Collaboration in largegroups (roughly more than 25) is enabled by stigmergy.Heylighen [11] proposed to distinguish stigmergy in theOSS community as direct and indirect. In OSS development,the unfinished jobs serve as the direct stigmergy, whichstimulates other actors to participate in finishing the jobs.Indirect stigmergy can be recognized in forums where bugsor function requests are posted. These forums are regularlyconsulted by the developers, thus attracting their attention totasks that seem worth performing. However, the numericalresearch and the mathematical stigmergy model are notdiscussed in Elliott’s and Heylighen’s publications. In thisresearch, we propose a stigmergy collaboration OSS modelto produce a simulation that accurately represents thecollaboration in an OSS community. The simulation outputsare compared with the empirical data retrieved from actualOSS project log information.IV. AGENT BASED SIMULATION MODELOur approach to reproducing the complex environment ofthe OSS software development community was to developan agent-based simulation framework using the stigmergyapproach. We assume that OSS members already haveenough motivation to join the community and makecontributions. The model will represent how the OSScommunity collaborates and how each individual developerchooses which forum to join. Our hypothesis is that thecollaborations of individual OSS developers and users arestigmergy collaborations. The forum posts and email listmessages serve as the digital stigmergy. Peer-to-peercommunications between individual OSS members do notoccur very often.To simplify the simulation, we assume there are two kindsof agents in the simulation - the developer and the user. The

developer agents voluntarily contribute their effort and timein answering questions from users, developing code andfixing bugs. The agents do not interact with each otherdirectly. Instead, they go through the forums for informationexchange. The user agents post messages on the forum. Auser agent can change to a developer agent if they want to.There are two kinds of forums, the public forum and thedeveloper forum. The public forum can be accessed andmessages can be posted by any agents interested in thesoftware project. Most of the time, it serves as the indirectmessage exchange board between users and developers.Users can post questions about how to use the software,bugs they found during software use and functions they wishto be included in the software. Each problem will berepresented by one forum thread. The other users and thedevelopers occasionally go through the forums, answer thequestions and get first hand information about bug problemsand wish list functionality.In this simulation, developers use the Ant ColonyAlgorithm to choose which forum thread problem tocontribute to solving. In this algorithm, each forum threadserves as one potential digital trail to different softwaredevelopment directions and the posted messages in thisthread represent the digital pheromones laid down on thetrail. Every time a user or developer posts a new message inthis forum thread, a new pheromone is deposited on the trail.The pheromone content of a forum thread can be updatedand decayed by using the following models.Pheromone update: when a message is posted in a forumthread, the pheromone for this thread is incremented by aconstant, γ. The nominal value of γ is one. Equation (1)describes the pheromone update procedure when a messageis posed by actor a in a post thread d at time t.ttP 1 d Pd (1)Pheromone decay: to account for pheromone decay, eachthread’s pheromone values are periodically multiplied by thedecay factor, ε -τ . The decay rate is τ0. A high decay ratewill quickly reduce the amount of remaining pheromone,while a low pheromone decay rate will degrade thepheromone value slowly. The nominal pheromone decayinterval (or decay period) is one day. Equation (2) describespheromone decay.t 1 tP * dPdIf no message has been posted in a thread in quite sometime, the pheromone for this thread will be decayed to anear-zero value. The thread will be removed from thedeveloper’s potential thread selection direction.ACO algorithm: The developers and users follow the ACO(Ant Colony Optimization) algorithm [12] in swarmintelligence for choosing the thread they join and post(2)messages. According to the pheromone theory, they willmost likely join and post messages in the thread that hashighest pheromone content.Thread selection: Actors will randomly chose a threadbased on the amount of pheromone present on each forumthread. Equation (3) describes thread d ’s probability dof being chosen.d N( Pi1td( P K)tiF K)N is the total number of forum threads. The constants Fand K are used to tune the forum-choosing behavior of anactor. The value of K determines the sensitivity of theprobability calculations to small amounts of pheromone. If Kis large, then large amounts of pheromone will have to bepresent before an appreciable effect will be seen in themessage posting probability. The nominal value of K is zero.Similarly, F may be used to modulate the differencesbetween pheromone amounts. For example, F > 1 willaccentuate differences between links, while F < 1 willdeemphasize them. F = 1 yields a simple normalization. Thenominal value of F is two.Another forum in the simulation is the developer forum. Itserves as the internal forum and is used by developers topost their ongoing work. It represents the email-list andSVN repository in an actual OSS project. Each developer’scontribution is stimulated by other developers’ ongoingwork posted in the developer forum. The probability c fordeveloper i to continually contribute to a software elementdevelopment is modeled as the termite mud drop probabilityin Grasse stigmergy model and is given by:cif ( i) ni*(k f ( i)jLF2)1 ( i,j)f ( i) max(0.0, (1 )2 Here, ( i,j)[0,1]is the dissimilarity value betweenideveloper i contributed post and the developerj contributed post message in email-list and SVNrepository, n is the message length. [0,1]is a data-2dependent scaling parameter, and is the total number ofposted messages in the developer forum during a predefinedtime period L [1,15].(3)(4)(5)

V. EXPERIMENTAL RESULTSWe used detailed OSS community log data onSourceForge to illustrate the model’s theoreticalmechanisms. SourceForge, an online center for OSSdevelopment communities, provides collaborative resourcesfor approximately 200,000 projects. This data consists of allthe activity information of OSS software developers andusers registered on SourceForge from 2003 to 2008. Wedeveloped scripts that query the SourceForge Research DataArchive hosted by the University of Notre Dame [5] forproject empirical data that meet our criteria. We wereinterested in OSS projects that reached a minimum team sizeof 20 developers at some point in the project lifetime inorder to insure that the social and technical factors were wellrepresented. In addition, we eliminated data on projects thatappeared to use SourceForge only as a means to downloadthe software, not as a development environment. In all, weidentified 67 projects as viable for use as data for our OSSmodel.Our first experiment simulates the forum postingbehaviours of users and developers in the OSS community.We use the values observed in empirical data derived fromreal log data in OSS repositories to calibrate and evaluateour stigmergy collaboration model. In our simulation, thereare 20 developers and 1000 users. The users randomly postnew message threads or leave messages on existing threadson the message board and leave pheromones on the messageboard. The developers respond to answer the questionsposted by users and choosing work that is posted on themessage board, such as bug fixes and new function modules.Both user and developer agents follow the stigmergy rulesdiscussed in the previous section. The simulation output isshown in figure 1. In figure 1, the x axis represents the logvalue of the number of total posted messages of each threadcreated by agents. The y axis represents the log value of thethread number in the message board. It indicates that a largenumber of threads have only a very small amount of postedmessages. But some threads have large numbers of postedmessages. The large amount of messages in some threadsnormally indicates there are some bugs or new softwarefunctions that need to be fixed or implemented. Thedeveloper agents who post messages in the thread will bemotivated to contribute to solving the problems discussed inthe thread. Developer agents will turn to the developerforum to start a new thread or to contribute work to anexisting thread in the developer forum. Because of spacelimitations, the developer forum simulation will be discussedin another publication.We compared the simulated outputs to empiricallyobserved behavior. Figure 2 represents the real averagedistribution of the thread and the message numbers for eachthread. The distribution patterns in figure 1 and figure 2indicate that the simulation is able to partially reproduce theforum evolution trend in many OSS projects. The closenessof the simulated results to the empirical data indicates thatour model may reflect the processes that occur in OSSevolution.Figure 1: Posted message number and thread number distribution –simulationFigure 2: Posted message number and thread number average distribution -OSS community dataVI. CONCLUSIONIn this research, we presented a stigmergy approach tosimulate the collaboration model in the OSS developercommunity. In this research, we examined how the selforganizingbehavior of individuals affects the emergence ofdifferent types of organizational characteristics. Thestigmergy concept provides a theory for explaining thecollaboration model in the OSS community and other onlinecommunities. Our hypothesis is that the collaborations ofindividual OSS developers and users are stigmergycollaborations. We compared the simulated output toempirically observed behaviors in different OSS projectforums. Our preliminary experiment simulation is able topartially reproduce the forum evolution trend and forumposting behaviors of OSS users and developers in manyOSS projects. The closeness of the simulated results to theempirical data indicates that our model may reflect the realprocesses that occur in OSS evolution. A thoroughvalidation needs to be conducted to prove the model’saccuracy. Our next research step will be to build asimulation to simulate the work choosing behaviors ofdevelopers and continue to investigate how the choicesagents make in contributing their work on projectsdetermines the global status of the whole OSS project.Further research will be to extend the research model toother self-organized communities. Our hypothesis is if thesame mathematical model captures the behavior of different

systems, then we can talk about similarities between systemsbeyond simple analogies. We can use our mathematicalmodels to make predictions that apply to many differentsystems. Furthermore, if two systems obey the samemathematical laws, we can perform experiments on onesystem and infer how another system might behave undersimilar conditions.ACKNOWLEDGMENTPrepared by Oak Ridge National Laboratory, P.O. Box2008, Oak Ridge, Tennessee 37831-6285, managed by UT-Battelle, LLC, for the U.S. Department of Energy undercontract DE-AC05-00OR22725; and by Lockheed MartinCompany. The views and conclusions contained in thisdocument are those of the authors and should not beinterpreted as representing the official policies, eitherexpressed or implied, of the Oak Ridge National Laboratory,Lockheed Martin Company, the Department of Energy orthe U.S. government.REFERENCES[1] Audris, M., T.F. Roy, and D.H. James, Two case studies of opensource software development: Apache and Mozilla. ACM Trans.Softw. Eng. Methodol., 2002. 11(3): p. 309-346.[2] Dorigo, M., E. Bonabeau, and G. Theraulaz, Ant algorithms andstigmergy. Future Generation Computer Systems, 2000. 16(8): p. 851-871.[3] Yutaka, Y., et al., Collaboration with Lean Media: how open-sourcesoftware succeeds, in Proceedings of the 2000 ACM Conference onComputer Supported Cooperative Work. 2000, ACM: Philadelphia,Pennsylvania, United States.[4] Xu, J., Y. Gao, and G. Madey. A Docking Experiment: Swarm andRepast for Social Network Modeling. in Seventh Annual SwarmResearchers Meeting (Swarm2003). 2003. Notre Dame, IN.[5] Christley, S. and G. Madey, Collection of Activity Data forSourceForge Projects. 2005, Dept. of Computer Science andEngineering, University of Notre Dame: Notre Dame, IN.[6] Crowston, K., et al. Towards a Portfolio of FLOSS Project SuccessMeasures. in Open Source Workshop of the International Conferenceon Software Engineering (ICSE 2004). 2004.[7] Hackman, J.R., The Handbook of Organizational Behavior, in Thedesign of work teams, J.W. Lorsch, Editor. 1986, Prentice-Hall:Englewood Cliffs, NJ.[8] Neil Smith, A.C., Juan Fernández-Ramil, Users and Developers: AnAgent-Based Simulation of Open Source Software Evolution, inSPW/ProSim 2006. 2006.[9] Jin Xu, Y.G., Scott Christley, Gregory R. Madey, A TopologicalAnalysis of the Open Souce Software Development Community inHICSS 2005. 2005.[10] Elliott, M., Stigmergic Collaboration: The Evolution of Group Work.M/C Journal, 2006. 9(2).[11] Heylighen, F., ed. Why is Open Access Development so Successful?Open Source Jahrbuch, ed. M.B.R.A.G. B. Lutterbeck. 2007,Lehmanns Media.[12] Dorigo, M., Ant colony optimization and swarm intelligence in 4thInternational Workshop, ANTS 2004, Brussels, Belgium, September 5-8, 2004 : proceedings. Lecture notes in computer science. 2004,Berlin: Springer.