Home

Coordination Tools for MAS Development and Deployment

1. Table 2 ReSpecT code for the basic implementation of the Contract Net Protocol 18 1 reaction out announcement Task current_time T current_agent Id out_r history time T new_task_announcement Task Id 2 reaction out bid Task MyBid current_time T current_agent Id out_r history time T new_bid_arrived Task Id MyBid 3 reaction in bids Task L pre current_time T out_r history time T task_announcement_expired Task 4 reaction out award Task TheBid current_time T out_r history time T contract_awarded Task TheBid Table 3 Debugging patterns tracing ReSpecT code to add tracing capability 1 reaction out ntasks 0 current_time T out_r start_time T 2 reaction out announcement Task in_r ntasks N Ni is N 1 out_r ntasks N1 3 reaction in task_rate _ pre rd_r ntasks N rd_r start_time TO current_time T1 DT is T1 TO Rate is N 1000 DT out_r task_rate Rate Table 4 Debugging patterns statistics ReSpecT code for computing the task announcement number and the task announcement rate 1 reaction out ntasks Task 0 out_r total_bids Task 0 2 reaction out announcement Task in_r ntasks Task N Ni is N 1 out_r ntasks Task N1 3 reaction out bid Task _ in_r total_bids Task N Ni is N 1 out_r total_bids Task N1 4 reaction in bids_per_task Tas
2. Selected Papers from AGP 00 Vol 48 of Electronic Notes in Theoretical Computer Science Elsevier Science B V pp 179 196 Omicini A amp Denti E 20016 From tuple spaces to tuple centres Science of Computer Programming 41 3 277 294 Omicini A amp Zambonelli F 1999 Coordination for Internet application development Jour nal of Autonomous Agents and Multi Agent Systems 2 3 251 269 Special Issue Coordi nation Mechanisms for Web Agents Rossi D Cabri G amp Denti E 2001 Tuple based technologies for coordination in A Omicini F Zambonelli M Klusch amp R Tolksdorf eds Coordination of Internet Agents Springer Verlag chapter 4 pp 83 109 Rover D Waheed A Mutka M amp Bakic A 1998 Software tools for complex distributed systems Toward integrated tool environments IEEE Concurrency 6 2 40 54 Schumacher M 2001 Objective Coordination in Multi Agent System Engineering Design and Implementation Vol 2039 of LNAI Springer Verlag Smith R G 1979 The contract net protocol High level communication and control in a dis tributed problem solver in Proceedings of the 1st International Conference on Distributed Computing Systems IEEE Computer Society Washington D C pp 186 192 Van Liederkerke M amp Avouris N 1995 Debugging multi agent systems Information and Software Technology 37 2 103 112 27 Wegner P 19
3. a manager requests the bid list reaction in bids Task pre in_r announcement Task out_r collect_bids Task reaction in bids Task L post in_r contract_state Task announced out_r contract_ state Task evaluating L reaction out_r collect bids Task L in_r collect_bids Task L in_r bid Task Bid out _r collect bids Task Bid L reaction out_r collect bids Task L in_r collect_bids Task L no_r bid Task out_r bids Task L sates the winner a manager 2 2 reaction out award Task Bid in_r contract_state Task evaluating L name out_r contract state Task awarded Bid node out_r provide responses Task L reaction out award Task Bid in_r award Task Bid heiper reactions o provide the responses reaction out_r provide responses Task Bid rd_r contract_ state Task awarded Bid out_r bid response Task Bid awarded reaction out_r provide responses Task Bid _ no_r contract_state Task awarded Bid out_r bid response Task Bid not_awarded reaction out_r provide responses Task List out_r provide responses Task List reaction out_r provide responses T L in_r provide responses T L reaction in bid response Task awarded post in r contract state Task Figure 8 The Inspector tool connected
4. A TuCSoN node is started by issuing the command java cp tucson jar alice tucson runtime Node Of course the path to the tucson jar archive can be added to the CLASSPATH environment variable instead of using the cp option Here the TuCSoN node also reports the creation of tuple centre temperature occurred as a consequence of the operation temperature out data jan 15 17 5 14 discussed in the text r r F TuCSoN Command Line Interpreter 0909 TuCSoN Command Line Interpreter command command output output ok ic 900 usr bin login ttyp5 Welcome to Darwin alice node 188 aricci java cp tucson jar alice tucson ide CLIAgent console TuCSoN Command Line Interpreter Version 1 2 0 Tue Jun 18 21 16 48 CEST 2002 gt Figure 2 The TuCSoN CLI Agent start up top left and issuing commands top right To start up a CLI Agent in TuCSoN 1 2 0 one just types the command java cp tucson jar alice tucson ide CLIAgent the console option launches the con sole non graphic version bottom instead of the standard graphic version 10 e C8 TuCSoN Command Line Interpreter command temperature set spec reaction out X out r backup X TuCSoN Command Line Interpreter command temperature rd backup output ok waiting for tuple backup X answer backup array index 1 value tomato Figure 3 Re programmin
5. Now interaction starts By means of a CLIAgent a manager first makes a new announce ment its effect on the tuple centre can be observed by inspecting the triggered reaction view for the dynamics and the tuple view for the static final effect Figure 9 In particular the out communication action of the Manager triggers the first reaction of Table 2 whose successful execution is reported in the triggered reaction view consequently the tuple view reports the new tuple inserted by the Manager and the tuple generated by the reaction Figure 10 shows the interaction steps of a bidder First the bidder reads the announcement top then issues its bid the tuple bid purchase_book isbn 88 06 15304 8 price 45 where price 45 represents the bid middle finally asks for a response bottom Figure 11 shows the state of both the tuple set and the pending query set after two more bidders have submitted a bid and are waiting for their answers the tuple view shows the details about the submitted bids bids with price 45 price 50 and price 47 while the pending query view reports the requests of the three bidders waiting for an answer A critical point in the protocol dynamics which is worth observing is the closing of the bid ding stage from the Manager Figure 12 top shows the Manager s CLIAgent requesting a tuple 14 of template bids purchase_book isbn 88 06 15304 8 BidList and receiving the tuple bids pu
6. bid_response purchase _book isbn 88 06 15304 8 price 47 not_awarded bid_response purchase_book isbn 88 06 15304 8 price 50 not_awarded The effect of the Manager communication act on bidders is shown in Figure 14 two bidders receive a negative response one receives the positive awarded response 4 3 Debugging patterns The programmability of the coordination medium makes it possible to specify in principle both general purpose and specific debugging patterns as coordination laws defining the medium behaviour To this end the expressiveness of the specification language is the key to enable and promote a fundamental synergy between the conceptual tools provided by the coordination model to support the specification of the debugging patterns and the deployment tools enacting and effectively supporting such patterns at run time In our case the expressiveness of the ReSpecT language makes it possible to specify any kind of debugging patterns which can then be adapted dynamically by need Moreover the deployment tools provided such as the Inspector specifically exploit the synergy with the conceptual tools suggested by the ReSpecT coordination model With reference to out case study let us consider the three following debugging patterns e Tracing Protocol evolution can be explicitly documented tracing each relevant step and easily observing the full history of the protocol interactions For instance the react
7. level observer This does not mean that all interferences can be prevented at all however their impact is reduced since they can occur only at the meta level 2 2 Mediated Interaction In mediated interaction models interaction never occurs directly between interacting entities communication occurs via some kind of communication abstraction explicitly introduced for the purpose of enabling communication A typical form of mediated interaction is blackboard based interaction Englemore amp Morgan 1988 there interacting entities write read messages to from the blackboard only This approach makes it possible to achieve some fundamental properties such as name uncoupling space uncoupling and time uncoupling that is interacting entities do not necessarily have to know each other explicitly nor to be in the same place not even to co exist at the same time in order to communicate Moreover the set of messages in the blackboard represents at any time the state of communication which can therefore be observed quite simply In tuple based interaction Rossi Cabri amp Denti 2001 originated by the Linda model Gel ernter amp Carriero 1992 the communication medium is the tuple space that is a multi set of tuples which are ordered collections of data chunks Agents still communicate by adding out removing in and reading rd tuples to from the tuple space but access is associative based on some matching mechanism like patt
8. to the contracts tuple centre and the specification view with the ReSpecT code of Table 2 13 4 1 Agents Behaviour and Coordination laws in CNP The first act of a Manager Table 1 left is to announce a task to be performed line 1 and then wait for bids When the related timeout expires the Manager retrieves the list of bids from potential contractors line 3 selects the best bid according to its evaluation criteria line 4 awards the contract line 5 and receives the result of the performed task line 6 On the other side a Contractor Table 1 right waiting for possibly interesting task an nouncements line 1 first evaluates its capability to respond to the current request line 2 then issues its bid line 3 and starts waiting for an answer line 4 Only if the bid is accepted by the Manager answer awarded the Contractor performs the task line 5 and outputs the computed result line 6 The coordination rules for basic CNP are expressed by the ReSpecT code in Table 2 Note that interaction acts performed by agents via in inp out rd rdp occur separately from the coordination level of tuple centre reactions based on in _r out _r verb rd_r no _r that specify coordination rules as algorithms operating on the tuple set The contract_state auxiliary tuple is used to track the protocol state during the coordina tion activity At the beginning when a Manager issues a task announcement the state is set to annou
9. 96 Coordination as constrained interaction in P Ciancarini amp C Hankin eds Proceedings of the Ist International Conference on Coordination Languages and Models Vol 1061 of LNCS Springer Verlag pp 28 33 Wegner P 1997 Why interaction is more powerful than computing Communications of the ACM 40 5 80 91 28
10. Coordination Tools for MAS Development and Deployment Enrico Denti DEIS Universita di Bologna Viale Risorgimento 2 40136 Bologna Italy http lia deis unibo it ed mailto edenti deis unibo it Andrea Omicini DEIS Universita di Bologna Via Rasi e Spinelli 176 47023 Cesena FC Italy http lia deis unibo it ao mailto aomicini deis unibo it Alessandro Ricci DEIS Universita di Bologna Via Rasi e Spinelli 176 47023 Cesena FC Italy http lia deis unibo it ar mailto aricci deis unibo it Abstract One of the most critical issues in the engineering of multiagent systems MAS is the inadequacy of the available tools for MAS development and deployment As we assume in teraction as a first class issue in MAS tools are particularly required to monitor and debug inter agents aspects such as interaction protocols coordination policies social norms and environment constraints Since we claim that the definition of such tools is a basic research issue in this paper we aim at identifying the main requirements for development and deploy ment tools within an effective agent infrastructure Focusing on agent interaction aspects we take tuple based coordination infrastructures in particular the TuCSoN technology and tools as our reference and discuss the role of tools in a simple case study the development and deployment of a well known agent interaction protocol the Contract Net 1 Coordination Issues in Agent based Deve
11. Omicini amp Denti 20016 are tuple spaces whose behaviour can be defined by means of reactions to communication events expressed in the Turing equivalent ReSpecT language Omicini amp Denti 2001a a tuple centre can then be programmed so as to feature potentially any desired behaviour Tuple centre behaviour is expressed in terms of ReSpecT specification tuples correspondingly a tuple centre can be thought as made of two parts the tuple space holding ordinary data tuples and the specification space holding specification tuples So both the state of communication and the state of coordination expressed uniformly as logic tuples are fully observable and modifiable Ad hoc primitives are also provided to add consult and remove specification tuples from the specification space Formally given the set S of specification tuples a tuple centre state is described at any time by the set T of ordinary tuples the set W of the currently pending queries and the set Z of triggered reactions waiting to be executed In particular if S is empty the tuple centre behaviour defaults to the standard tuple space behaviour These sets provide different relevant views over interaction The current communication state is captured by T and W more precisely T being concerned with the data exchanged among agents provides a data oriented view over communication whereas W being concerned with the act of communication provides a control oriented view ov
12. al reaction in bids Task L pre current _time T out_r history time T task_announcement_expired Task reaction out award Task TheBid current_time T out_r history time T contract_awarded Task TheBid awara items 6 history time 6312267 contract_awarded purchase book isbn 88 06 15304 8 price 45 history time 6301922 new_task_announcement purchase _book isbn 88 06 15304 8 user 1024350268708 137 204 191 189 history time 6312037 task_announcement expired purchase book isbn 88 06 15304 8 history time 6302383 new bid arrived purchase _book isbn 88 06 15304 8 user 1024350268648 137 204 191 189 price 45 history time 6302192 new_bid arrived purchase book isbn 88 06 15304 8 user 1024350265153 137 204 191 189 price 47 history time 6302253 new_bid arrived purchase book isbn 88 06 15304 8 user 1024350268738 137 204 191 189 price 50 pe get any new observation get only when update requested Figure 15 The specification view with the new reactions added for tracing the main interaction event top and the tuple set view after a run of the CNP with the extended specification bottom 26 Denti E amp Omicini A 1999a An architecture for tuple based coordination of multi agent systems Software Practice amp Experience 29 12 1103 1121 Denti E amp Omicini A 19996 E
13. and the related development and deployment tools CLIAgent and Inspector have been used as a reference for the discussion and for the development of the Contract Net Protocol case study Future work includes the improvement of some services provided by In spector tool in particular the inspection of the dynamics inside the tuple centre virtual machine from the observation of triggered reactions to the management of the virtual machine control flow We expect to develop a more flexible presentation of the triggered reactions providing different levels of detail about reaction information and filtering options For the second issue we expect to support the step by step tracing of other fine grained interaction events such as an imminent listening speaking transition as well as the interception of all the events described in tuple centre operational semantics Omicini amp Denti 2001b We also plan to trace the com plete execution of ReSpecT programs showing the step by step execution of single reactions or ReSpecT operation such as out_r in principle we could even provide support for undoing them More generally what to do on reactions at run time from tracing and single step exe cution to breakpoints on reactions dynamic addition removal of reactions transparently to the ReSpecT virtual machine i e operating at the meta level is a matter for further exploration Finally one of the main challenge for the future is to devise
14. anguages and architectures provide abstractions and tools to manage and rule the agent interaction space Effectively governing interaction basically means the ability to specify and possibly enforce goals constraints and desired properties that are not peculiar to a given agent but to an ensemble of agents and also the ability to adapt modify this social glue dynami cally at run time all without affecting agent autonomy While computation languages express the inner working algorithm of an agent coordination languages express the agent s observable behaviour what is needed to design its interaction protocol Dependencies in MAS can be classified in two classes Schumacher 2001 objective and sub jective basically according to the viewpoint over the space of agent interaction that is adopted Objective dependencies refer to inter agent dependencies and typically concern the configu ration of the system in terms of the basic interaction media agent generation destruction and environment organisation subjective dependencies instead refer to the intra agent depen dencies toward other agents The management of subjective dependencies is called subjective coordination and is essentially concerned with intra agent aspects The management of objec tive dependencies instead is called objective coordination and is essentially concerned with inter agent aspects since the dependencies are external to agents So in the subjecti
15. book isbn 88 06 15304 8 price 47 price 45 price 50 no_r contract_state purchase book isbn 88 06 15304 8 awarded price 47 out_r bid response purchase book isbn 88 06 15304 8 price 47 not_awarded OK time 1024347646578 reaction out_r provide responses purchase book isbn 88 06 15304 8 price 45 price 50 in_r provide_ responses purchase_book isbn 88 06 15304 8 price 45 price 50 OK time 1024347646578 reaction out_r provide responses purchase book isbn 88 06 15304 8 price 50 no_r contract state purchase book isbn 88 06 15304 8 awarded price 50 out_r bid response purchase book isbn 88 06 15304 8 price 50 not_awarded OK Figure 13 The CLIAgent of the Manager communicating the awarded bid and the part of the triggered reactions in the related view 24 waiting for tuple bid response purchase book isbn 88 06 15304 8 price 47 Answer answer bid response purchase_book isbn 88 06 15304 8 price 47 not_awarded TuCSoN Command Line Interpreter command contracts s206 ingce unibo it in bid response purchase book isbn 88 06 15304 8 price 45 Answer waiting for tuple announcement Task answer announcement purchase book isbn 88 06 15304 8 ok waiting for tuple bid response purchase _book isbn 88 06 15304 8 price 45 Answer ja
16. d connection problem i e how to find the appropriate agent for a given task According to the terminology defined by Huhns amp Stephens 1999 Manager agents have tasks to be solved while Contractor agents are able to solve tasks The involved coordination aspects are expressed in terms of coordination laws by the ReSpecT code shown in Table 2 while the behaviour of both Managers and Contractors is described by the pseudo code in Table 1 agent communication acts are put in evidence It should be clear that such a code implements a very basic CNP version whose purpose is just to show how coordination occurs not to actually constitute a full fledged CNP implementation In particular we assume here that the protocol is used in a purely cooperative context accordingly the protocol is designed to minimise interactions not to prevent interferences from malevolent competitive agents G oes usr bin login ttyp1 Welcome to Darwin alice node 188 aricci java cp tucson jar alice tucson runtime Node TuCSoN Node Management Agent booted TuCSoN Infrastructure version 1 2 8 installed on port 20504 Wed Jun 19 69 22 13 CEST 2682 TuCSoN Node checking context default TuCSoN Node create new tuple centre temperature TuCSoN Node accepting user user 1624471423227 137 264 191 189 to default context TuCSoN default context user user 1624471423227 137 264 191 189 enters the context gt Figure 1
17. e book isbn 88 06 15304 8 price 47 contract state purchase book isbn 88 06 15304 8 announced announcement purchase book isbn 88 06 15304 8 get only when update requested 4 were iii SSS items 3 from agent user 1024345286463 1 local tim in bid_ response purchase book isbn 88 06 15304 8 price 45 Answer in bid_ response purchase _book isbn 88 06 15304 8 price 50 Answer from agent user 1024345900960 1 in bid_ response purchase book isbn 88 06 15304 8 price 47 Answer from agent user 1024346238551 1 get any new observation get only when update requested Figure 11 The tuple set view and the pending query view after the proposals of three bidders waiting for the answer 22 TuCSoN Command Line Interpreter waiting for tuple bids purchase_book isbn 88 06 15304 8 BidList bids purchase book isbn 88 06 15304 8 price 47 price 45 price 50 reaction out_r collect_bids purchase _book isbn 88 06 15304 8 price 50 in _r collect _bids purchase book isbn 88 06 15304 8 price 50 in r bid purchase book isbn 88 06 15304 8 _ out_r collect_bids purchase_book isbn 88 06 15304 8 _ price 50 OK time 1024346768946 reaction out_r collect_bids purchase _book isbn 88 06 15304 8 price 45 price 50 in_r collect_bids purchase book i
18. e default tuple centre of the local node If everything goes well the ok message will be displayed in reply To read back the above tuple an agent either the same or another one should issue a rd command whose template matches the inserted tuple like the following rd p X Y The infrastructure will then reply with a response message providing the bindings for the X and Y variables in this case hopefully something like answer p 1 hello To interact with other tuple centres than the default one all is needed is to specify the desired tuple centre name in the command invocation For instance to insert some data in the temperature tuple centre one just issues the command Figure 2 right temperature out data jan 15 17 5 14 The effect of this command is to place the tuple data jan 15 17 5 14 in the required tuple centre If it does not exist yet it is created on the fly as if it had existed since ever to the agents eyes an information message is shown in the TuCSoN node console Figure 1 Rather obviously on the fly creation is performed transparently the first time a new tuple centre name is referenced independently of the communication primitive which contains such a reference All the above commands apply to the LuCe infrastructure too Instead the network awareness feature is peculiar to TuCSON so in the TuCSoN context the coordination primitives can also address remote tuple centres identified by
19. er communication The current coordination state is instead represented by S and Z S contaning the static specification of the coordination rules provides a static view over coordination whereas Z along with the execution step of the tuple centre For details on tuple centre and ReSpecT semantics we forward the interested reader to Omicini amp Denti 20016 and Omicini amp Denti 2001a respectively virtual machine represents the dynamic enactment of coordination laws and provides a dynamic view over coordination 2 4 Deployment with Mediation When interaction is mediated by some kind of explicit communication media observing inter action no longer requires changes to the interacting entities since the communication state is represented and observable in the communication medium This chance does not mean however that any interference with the system is automatically prevented in standard tuple spaces for instance the insertion of a given tuple can only be observed by explicitly checking this kind of event by having an agent performing a suspensive matching rd So observation still takes place at the same level where communication occurs this holds even if the whole tuple set is retrieved via a sort of rdAl1 or copyAll1 bulk primitive which provides a snapshot at that given time of the tuple space state So the interaction histories of this communication medium change with respect to the case where the observer agent i
20. ern matching or unification The existence of explicit communication media raises interaction to a first class issue from the model to the implementation interaction is promoted to deserve explicit abstractions and places blackboards or tuple spaces devoted to it and to its own handling This key difference with respect to the previous approaches is what makes the communication state not only explic itly available but also observable without affecting the communicating entities in fact since all communication acts occur in given place s putting such places under observation requires no changes on the original system configuration avoiding the probe problem at that level a priori 2 3 Coordination and Tuple Centres While communication means enabling interaction coordination means constraining and ruling interaction Wegner 1996 In Linda coordination laws are built in the tuple spaces since the primitive operations and their semantics are pre defined and fixed If such laws are inadequate to the application changing the tuple space behaviour is impossible the desired overall behaviour must be built upon the fixed tuple space behaviour charging agents to fill the gap thus violating the separation between computation and coordination To overcome this limitation one must either add new primitives to the basic tuple space model or change the semantics of the existing operations as needed Rossi et al 2001 Tuple centres
21. g of interaction In order to be effective such methodologies and tools should fully support and promote the coordination model endorsed by the infrastructure in particular the tools for system develop ment and deployment should be inspired by and modelled after the metaphors and abstractions of the coordination model otherwise the gap between theory and practice would likely make the design development and deploying task basically unmanageable for any non trivial system We claim that the definition of such methodologies and tools is a fundamental research issue it is the researchers responsibility to analyse and indicate how the coordination model s concepts and run time model should transpose onto suitable methodologies and tools so as to ensure fundamental properties for system engineering incremental development system verification debugging and maintenance dynamic configurability etc 2 Deployment Tools for Interaction Here we focus on run time environments and tools for the support of deployment from de bugging to testing from monitoring to analysis and run time verification of the interaction related aspects of MAS While literature on tools and toolkits for the design and development of agent systems and agent infrastructures abounds very little seems to be available concern ing deployment tools Ndumu Nwana Lee amp Collis 1999 Perhaps quoting Van Liederkerke amp Avouris 1995 this is due to the fact
22. g the tuple centre behaviour reaction top and effect after an out bottom 7 Logic Tuple Centre Inspector tuple centre information name node 09c TuCSoN inspector spawner mtuple centre to inspect name default node localhost Figure 4 The Inspector Spawner left is launched issuing the command java cp tucson jar alice tucson ide InspectorSpawner Then the Inspector tool window appears right Al ternatively the Inspector can be activated directly by the command javaw cp tucson jar alice tucson ide Inspector TCName TCAddress Since TCAddress can be any Internet ad dress Inspectors can be launched on any remote node if TCAddress is omitted localhost is assumed Here too the Inspector main window eventually appears 11 eee Tuple Set of default localhost freee Tuple Set of default localhost vm time 1024426363451 local time 1024426363657 items 0 vm time 1024426545597 focal time 1024426545674 items 3 Enforce tuple set eee Tuple Set of default localhost reee Tuple Set of default localhost vm time 1024426669084 local time 1024426669283 items 2 local time 1024426742891 jtems 3 p ciao pl Figure 5 The Inspector tuple view observation mode top left action on the tuple set top right filtered view of the tuple set bottom left and log filtered view botto
23. ion block shown in Table 3 enables tracing i the announcement of a new task reaction 1 ii the occurrence of a new bid reaction 2 iii the announcement expiration reaction 3 and iv the contract award reaction 4 Each of these interaction events triggers a reaction that reifies the event occurrence by adding to the tuple set a tuple describing the event along with the event time stamp and the responsible agent In Figure 15 the top figure shows the specification view after updating the ReSpecT source with the new reactions tuples providing for tracing while the bottom figure shows the tuple set view after a complete run of the CNP with the new coordination specification containing history tuples about task announcement bid submission and contract award 15 e Statistics The possibility of creating and real time updating statistics on the interaction dynamics can be fundamental to improve the design of coordination activities in particular toward time scalability to understand how to adapt the coordination laws to minimise interactions and space scalability to understand how to design the topology of the coordination space number and distribution of tuple centres For instance the reaction block in Table 4 adds the capability to monitor the total number of announced tasks and the task rate contracts per second handled by a tuple centre this information provides one first yet coarse estimation of the coordinat
24. ion burden workload of the tuple centre As a further example the reaction block in Table 5 adds the capability to trace the mean number of bids for a specific task which can help to perform a deeper fine grained analysis of the coordination burden e Monitoring logic events The capability to monitor the dynamics of interaction protocols and notify the occurrence of selected logic events is a very effective tool for debugging and testing interaction in particular to detect logic errors in the coordination flow in the same way as assertions support the debugging and testing of algorithms As an example the reaction block in Table 6 adds the capability to monitor the partial number of bids for each run of a specific task announcement whenever such a number exceeds the threshold value stored in the monitor_bids Task Threshold NotificationCentre tuple which is obviously supposed to be present in the tuple set a warning tuple like bids_overflow SourceTC Task Threshold N is inserted in the NotificationCentre tuple centre by an out_tc operation thus reifying the event reaction 4 One first point worth remarking is that multi agent systems that do not have interaction media as first class entities usually provide some of these debugging features as a closed set of hard coded mechanisms typically bounded to a specific implementation platform Here instead infinite debugging strategies can be provided as programs in a Turing equivalent i
25. k NBids pre rd_r total_bids Task NBids rd_r ntasks Task NTasks Value is NBids NTasks out_r bids_per_task Task Value Table 5 Debugging patterns statistics ReSpecT code for computing the mean value of bids per task 19 1 reaction out announcement Task out_r partial_bids Task 0 2 reaction in r announcement Task in_r partial_bids Task _ 3 reaction out bid Task _ in_r partial_bids Task N Ni is N 1 out_r partial_bids Task N1 4 reaction out_r partial_bids Task N rd_r monitor_bids Task Threshold NotificationCentre N gt Threshold current_tc SourceTC out_tc NotificationCentre bids_overflow SourceTC Task Threshold N Table 6 ReSpecT code adding logic event notification capability TuCSoN Command Line Interpreter command output lok ion Set of contracts s206 ingce unibo it time 1024345152201 reaction out announcement purchase book isbn 88 06 15304 8 out_r contract state purchase book isbn 88 06 15304 8 announced OK Fin t f i store i f E beg file inspector reactions log browse i 4 OOO Tuple Set of contracts s206 ingce unibo it vmtime 1024345152201 local time 10243450 contract state purchase book isbn 88 06 15304 8 announced announcement purchase _ book isbn 88 06 15304 8 f Observation View Log Action t
26. lopment A multiagent system MAS is more than just a sum of individual agents how agents interact with other agents and the environment and how they are organised in an ensemble cannot be reasonably handled as individual issues In this respect interaction should be considered a fundamental design dimension the dimension of the social aspects that determine collective behaviour Therefore interaction should not be treated as a mere afterthought following the design of the individual system components instead it is indeed a critical design issue to be accounted for since the earliest design phases But what does taking interaction into account actually mean Just defining and imple menting a set of interaction protocols whatever mechanism is used is not enough since this would typically not ensure that the agent ensemble behaves as desired respecting global constraints and ensuring the achievement of the desired global system Instead exploiting inter action as a crucial design dimension requires the ability to govern it in a precise well specified and controlled way Yet mastering interaction is much a less established area Wegner 1997 Constraining and governing interaction is precisely the purpose of coordination models and languages Gelernter amp Carriero 1992 in the same way as agent languages and architectures provide abstractions and tools to design and build the individual agents coordination mod els l
27. m right r Pending Query Set of default localhost ee e Behaviour of tuple centre default localhost reaction out p X in_r p X in_r q Y out_r pqiX vm time 1024427334568 local time 1024427334765 items jreaction out q in_r p x im_r q out_r pqiX rd p X from agent user 1024425949767 137 204 191 188 in q X 1 from agent user 1024426189180 137 204 191 188 pe get any new observation get only when update requested __ Triggered Reaction Set of default localhost time 1024427063334 reaction out p 4 in_r p 4 in_r q _ out_r pq 4 _ FAILED time 1024427175138 reaction out q 5 in_r p _ in_r q _ out_r pq _ 5 OK store a logfile inspector reactions log 4 Figure 6 The Inspector pending query view top left the specification set right and the triggered reactions view bottom left 12 Logic Tuple Centre inspector tuple centre information name node Figure 7 The Inspector tool support for step by step tracing of the tuple centre virtual machine T Behaviour of tuple centre contracts s206 ingce unibo it setup a contract with the task announcemen reaction out announcement Task out _r contract state Task announced uple centre to inspec name contracts node s206 ingce unibo it
28. nced reaction 1 Then when the Manager retrieves the list of proposals meaning that the bidding stage has finished the contract state is updated to evaluating which also specifies the retrieved proposal list reaction block 2 When the Manager eventually communicates the awarded bid the contract state is updated accordingly and a set of response tuples of the form bid_response Task Bid Answer are generated Answer is awarded in the winning bid and not_awarded in the others reaction block 3 Finally when the response is retrieved by the awarded bidder the last coordination law cleans up the contract state reaction 4 4 2 TuCSoN Tools for CNP debugging To show the impact of the TuCSON support tools let us consider the run time inspection of the tuple centre during the enactment of the Contract Net Protocol among a set of agents As a specific example we consider a CNP run where the task to be assigned is represented by the tuple purchase_book isbn 88 06 15304 8 while interacting agents are humans acting by means of the CLIAgent tool through the contracts tuple centre located at the TuCSoN node s206 ingce unibo it First we use an Inspector to set the coordination specification reported in Table 2 Figure 8 to this end we open a specification view write the source code or load it from a file or paste it via copy amp paste have it syntax checked and upload it to the tuple centre by means of the accept button
29. new metaphors specially suited to the management and debugging of interaction oriented virtual machines going beyond the classical metaphors traps breakpoints watches etc of the algorithmic world Acknowledgements First we would like to thank MIUR the Italian Ministero dell Istruzione dell Universita e della Ricerca and Nokia Research Center Burlington MA USA for partially supporting this work Also let us thank Paolo Petta and J P Miiller organisers of the Third International Symposium From Agent Theory to Agent Implementation AT2AI 3 and all the attendants as well which stimulated the further development of the concepts and ideas presented here References Bates P 1995 Debugging heterogeneous distributed systems using event based models of behaviour ACM Transaction on Computer Systems 13 1 1 31 Cremonini M Omicini A amp Zambonelli F 1999 Multi agent systems on the Internet Extending the scope of coordination towards security and topology in F J Garijo amp M Boman eds Multi Agent Systems Engineering Vol 1647 of LNAI Springer Verlag pp 77 88 17 1 out announcement Task 1 rd announcement Task 2 wait ExpireTime 2 MyBid evaluate Task 3 in bids Task BidList 3 out bid Task MyBid 4 Bid selectWinner BidList 4 in bid_response Task MyBid Answer 5 out award Task Bid 5 if Answer awarded 6 in res
30. ng debugging patterns In particular since any condition can in principle be intercepted by some suitable reaction and handled accordingly tuple centre based debugging suffers no a priori limitations This is a much more flexible approach than many commercial packages TupleScope by Scientific Computing s Linda implementation described in Linda user manual for instance allows only a pre defined set of conditions to be captured and monitored those that TupleScope s designers judged to deserve such special treatment We should finally remark however that observation problems while prevented at the com munication level obviously re appear at the meta level in both the static and the dynamic views In fact the set S of specification tuples can be observed statically through some ad hoc primitive indeed with the same limits discussed above for the tuple space showing that observation problems have just moved one level up From the dynamic viewpoint tuple centres enable the set Z of triggered both failed and successful reactions to be observed along with the current state of the tuple centre virtual machine as expected the probe problem is still there we can add reactions to help debugging the existing reactions but cautions have to be taken again in order to avoid interferences since all actions occur at the same meta level For details on event observability and coordination primitive semantics we forward
31. ngineering multi agent systems in LuCe in S Rochefort F Sadri amp F Toni eds ICLP 99 International Workshop on Multi Agent Systems in Logic Programming MAS 99 Las Cruces NM Englemore R S amp Morgan A eds 1988 Blackboard Systems Addison Wesley Garcia Molina H Germano F amp Kohler W 1989 Debugging a distributed computing system IEEE Transaction on Software Engineering 10 2 210 219 Gelernter D amp Carriero N 1992 Coordination languages and their significance Communi cations of the ACM 35 2 97 107 Graham J McHugh D Mersic M Foster M Windly V Cleaver D amp Decker K 2001 Tools for developing and monitoring agents in distributed muti agent systems in Proc of Autonomous Agents 2000 Workshop on Infrastructures for Scalable MASs Huhns M amp Stephens L 1999 Multiagent systems and societies of agents in G Weiss ed Multiagent Systems A Modern Approach to Distributed Artificial Intelligence The MIT Press pp 79 120 McDowell C amp Helmbold D 1989 Debugging concurrent programs ACM Computing Sur veys 21 4 593 622 Ndumu D Nwana H Lee L amp Collis J 1999 Visualizing and debugging distributed multi agent systems in Proceedings of Autonomous Agents 1999 ACM Omicini A amp Denti E 2001a Formal ReSpecT in A Dovier M C Meo amp A Omicini eds Declarative Programming
32. nswer bid response purchase_book isbn 88 06 15304 8 price 45 awarded command contracts s206 ingce unibo it in bid response purchase book isbn 88 06 15304 8 price 50 Answe waiting for tuple announcement Task answer announcement purchase book isbn 88 06 15304 8 ok waiting for tuple bid response purchase book isbn 88 06 15304 8 price 50 Answer answer bid response purchase book isbn 88 06 15304 8 price 50 not_awarded Figure 14 The CLIAgent of the three bidders receiving the answer to their proposal only one is awarded the middle one 25 T Behaviour of tuple centre contracts s206 ingce unibo it no_r contract_ state Task awarded Bid out_r bid response Task Bid not_awarded reaction out_r provide responses Task ail List out_r provide responses Task List reaction out_r provide responses T L in_r provide responses T L J ng the contract reaction in bid response Task awarded post in_r contract_state Task tracing task announcement reaction out announcement Task current _time T current_agent Id out_r history time T new_task_announcement Task Id tracing bid proposal reaction out bid Task MyBid current_time T current_agent Id out_r history time T new_bid arrived Task Id MyBid tracing bids retriev
33. nteraction oriented language like ReSpecT whose semantics is formally defined Moreover a suitable engineering of the interaction protocol makes it possible to develop a debugging policy incrementally as shown in the examples coordination laws here reactions realising debugging strategies can be progressively added to the core coordination rules without interferences So the encapsulation of the coordination specification provided by tuple centre is effective for engineering both the coordination activities and the debugging testing patterns 5 Lessons Learnt According to our exploration debugging is much easier and flexible when some form of mediated interaction blackboards tuple spaces tuple centres etc is adopted instead of point to point communication whether they are low level sockets or high level abstractions such as KQML or FIPA ACL Mediated interaction enables specific monitoring and debugging tools to be associ ated to the communication media directly thus bringing the metaphors used to model the agent interaction space down to the development and deployment phases The benefits are even more enhanced in coordination infrastructure based on tuple based models thanks to the uncoupling properties of generative communication and in the case of programmable coordination media such as the tuple centres to the ability of defining the behaviour of the communication media From the software engineering viewpoint we e
34. or the API available in both the Java and Prolog environments To test whether a TuCSoN node is running on a given host one can simply query the selected host by means of a telnet client for instance via a command like telnet localhost 20504 if present the TuCSoN run time as of version 1 2 0 will reply like this TuCSoN Infrastructure version 1 2 0 Current date Tue Jun 18 21 29 16 CEST 2002 Running since Tue Jun 18 21 21 21 CEST 2002 3 2 Tools In order to support the engineering of agent applications based on tuple centre coordination media TuCSoN provides suitable IDE tools specifically tailored to the abstractions endorsed by the coordination infrastructure the Command Line Interpreter Agent and the Inspector 3 2 1 Agent Tools The Command Line Interpreter Agent also called the CLI Agent tool enables users to interact with TuCSoN tuple centres anywhere on the network in the same way as software Java or Prolog agents would do A non graphic version is also available for devices with limited graphic support and platforms not supporting the Java Swing API see Figure 2 This tool lets users access tuple centres by means of Linda like communication acts i e coordination primitives such as out in rd inp rdp moreover they enable users to access the specification space of tuple centres via the set_spec and get_spec primitives For instance typing out p 1 hello asks the infrastructure to put the tuple p 1 hello in th
35. ous Omicini amp Denti 20016 an agent emitting a tuple has the ok message shown on its console immediately even if the out primitive itself is processed by the virtual machine only when the nezt button is pressed as it can be easily observed via the Inspector s tuple view 3 2 3 Some small examples The ability to specify the reactive behaviour of the coordination medium can be exploited to define custom observation and debugging policies In fact the Turing equivalence of ReSpecT provides the flexibility to express computations for tracing monitoring and debugging purposes which can be particularly useful when debugging complex multi stage interaction protocols involving different agent roles In these situations a tuple centre can be programmed so as to trace the protocol evolution making it easier to identify faults wrong malicious agents interaction behaviour etc For instance the following rule monitors the communication acts of agents providing a given information e g a product price and generates the related history reaction out price Info current_agent Id current _time Time out_r price_history time Time who Id content Info The rule below instead handles statistical information about service requests reaction in service Descr pre in_r service_counter Descr N N1 is N 1 out_r service_counter Descr N1 A familiar example is counting the number of hits of a
36. p X in r p X in r q Y out _r pq X Y reaction out q Y in r p X in_r q Y out_r pq X Y When p 4 is output the first reaction is triggered but fails due to the absence of any q Y tuple upon insertion of q 5 the second reaction is triggered and succeeds due to the presence of p 4 The result is that atomically both p 4 and q 5 are removed and the new tuple pq 4 5 is inserted In addition the Inspector enables the management of tuple centres as coordination virtual machines providing step by step tracing of the machine behaviour as traditional debuggers do This is done by selecting the virtual machine tab Figure 7 and enabling the step execution option from there tuple centre operations can be executed singly pressing the nezt button Each time the button is pressed the tuple centre virtual machine processes one communication event and all the reactions it triggers in a non deterministic order as single macro step of the virtual machine At the same time if desired the triggered reactions can be observed in the Inspector s triggered reaction view thus effectively supporting distributed remote debugging and monitoring If for instance an agent performs a rdp operation with the step execution option enabled the tuple centre will not reply to the agent immediately instead it will reply only when the next button is pressed Since the semantics of the out primitive is asynchron
37. rchase_book isbn 88 06 15304 8 price 47 price 45 price 50 as re sult the effect of the agent request on the tuple centre are shown in Figure 12 The triggered reaction view reports some of the reactions triggered by the request in particular the last reaction at the bottom of the text area is the successful reaction that inserts the tuple satisfying the manager request That reaction is a runtime instance of the last reaction of block 2 in Table 2 that succeeds when no more bids can be collected Figure 12 shows the final effect on the tuple set of the execution of both the initial request and the subsequent reactions no more bid tuples are present and the tuple set includes only the tuple below describing the new interaction state the evaluation stage contract_state purchase _book isbn 88 06 15304 8 evaluating price 47 price 45 price 50 Another critical point in the protocol dynamics is when the manager awards the contract after collecting the bids Figure 13 shows both the manager CLIAgent inserting the tuple award purchase_book isbn 88 06 15304 8 price 45 in the tuple centre and parts of the triggered reactions In particular the second the fifth and the seventh reaction belong ing to the reaction block 3 of the specification shown in Table 2 insert the following response tuples for the bidders bid_response purchase_book isbn 88 06 15304 8 price 45 awarded
38. re meta level with respect to the agent tuple centre interaction So the state of communication between that tuple centre and agents can be observed from the meta level without incurring in the probe problem Practically an inspector can be spawned in two ways either i by means of the Inspector Spawner application which provides a GUI to specify the name and the address of the tuple centre to be inspected or ii by providing such information directly at the command prompt Figure 4 The Inspector allows users to inspect 1 the set of tuples button tuples 2 the set of pending queries button pending 3 the set of behaviour specifications button specification 4 the set of triggered reactions button reactions The tuple view Figure 5 enables the dynamic observation of and the action on the set T of the logic tuples stored in the tuple centre see Subsection 2 3 The full content of the tuple set can be replaced in one shot by editing the list of displayed tuples in Figure 5 top right this feature is exploited to add some tuples to the tuple set Observation can be both real time and on demand in the first case each change occurring in the set is immediately reflected in the tuple view while in the second case a snapshot of the tuple set is taken only when explicitly requested via the Inspector s GUI Figure 5 top left It is also possible to filter the displayed tuples by specifying a t
39. s not present Also some communication events may not be observable by definition in standard tuple spaces rd operations cannot be observed since they cause no change to the communication state i e to the set of tuples Even the out operation dependently on the implementation may not be observable if an agent suspended on a matching in operation removes the tuple before it is actually added to the tuple space Though mediated interaction does not guarantee absence of interferences yet it is the key to reduce such interferences bounding them inside an upper meta level In tuple centres in particular both the communication and the coordination state are explicitly represented from the static and the dynamic viewpoints Since the tuple centre model is inherently able to perform operations reactions that are invisible to the interacting entities there is a degree of openness that makes it possible to add observation inside the communication media without affecting the communication level itself that is i no extra operations must be added in order to achieve observability and ii all communication events are fully observable whether they change the tuple space state or not This feature provides a fundamental new tool enabling us to monitor debug and test interac tion in a configurable way So it is possible to implement virtually any computable debugging policy possibly providing some frequent debugging rules as recurri
40. sbn 88 06 15304 8 price 45 price 50 in x bid purchase _book isbn 88 06 15304 8 _ out_r collect _bids purchase book isbn 88 06 15304 8 __ price 45 price 50 OK time 1024346768966 reaction out_r collect bids purchase book isbn 88 06 15304 8 price 47 price 45 price 50 in_r collect_bids purchase book isbn 88 06 15304 8 price 47 price 45 price 50 in r bid purchase _book isbn 88 06 15304 8 _ out_r collect _bids purchase _book isbn 88 06 15304 8 __ price 47 price 45 price 50 FAILED time 1024346768976 reaction out_r collect_ bids purchase _book isbn 88 06 15304 8 price 50 in_r collect_bids purchase_book isbn 88 06 15304 8 price 50 no r bid purchase _book isbn 88 06 15304 8 _ out_r bids purchase _book isbn 88 06 15304 8 price 50 FAILED time 1024346768976 reaction out_r collect_bids purchase_book isbn 88 06 15304 8 price 47 price 45 price 50 in_r collect_bids purchase_book isbn 88 06 15304 8 price 47 price 45 price 50 no_r bid purchase_book isbn 88 06 15304 8 _ out_r bids purchase_book isbn 88 06 15304 8 price 47 price 45 price 50 OK al logfile inspector reactions log pe get any new observation get only when update requested Figure 12 The CLIAgent of the manager reque
41. selected memory location or CPU registers others go further enabling traps on inter process communication the ability to add remove inter process messages and the chance of controlling timeouts McDowell amp Helmbold 1989 Garcia Molina et al 1989 The most recent and advanced debugging monitoring tools are event based Bates 1995 by catching event sequences they make it possible to post analyse and post process event histories so as to reconstruct the distributed system behaviour However it is known McDowell amp Helmbold 1989 that observing the execution of concur rent processes is not a neutral operation since it can interfere with the system behaviour probe effect This problem can be easily understood since the addition of any spy print or trace operation alters the timing of a process So for instance critical races which could be the cause of the system malfunctioning and the very reason for debugging may disappear in the new debug oriented system configuration Ndumu et al 1999 These problems are due to the lack of separation between computation and interaction the state of communication is not explicitly represented and so can not be observed without affecting the actors of communication which is precisely what changes irremediably their original behaviour As a general principle observing a given level with no strong interferences requires an explicit representation of that level and consequently a meta
42. sting the bid list and the related effect on the tuple centre triggered reactions and tuple set view 23 TuCSoN Command Line Interpreter command contracts s206 ingce unibo it out award purchase book isbn 88 06 15304 8 price 45 waiting for tuple bids purchase_book isbn 88 06 15304 8 BidList answer bids purchase book isbn 88 06 15304 8 price 47 price 45 price 50 ok Triggered Reaction Set of contracts s206 ingce unibo it reaction out_r provide responses purchase book isbn 88 06 15304 8 price 50 in_r provide_ responses purchase_book isbn 88 06 15304 8 price 50 OK time 1024347646558 reaction out_r provide responses purchase book isbn 88 06 15304 8 price 45 price 50 rd_r contract state purchase _book isbn 88 06 15304 8 awarded price 45 out_r bid_ response purchase book isbn 88 06 15304 8 price 45 awarded OK time 1024347646558 reaction out_r provide responses purchase book isbn 88 06 15304 8 price 4 price 45 price 50 in_r provide_ responses purchase_book isbn 88 06 15304 8 price 47 price 45 price 50 OK time 1024347646568 reaction out award purchase book isbn 88 06 15304 8 price 45 in_r award purchase book isbn 88 06 15304 8 price 45 OK time 1024347646568 reaction out_r provide responses purchase
43. that designing and building MAS is a complex pro cess the integrated system debugging phase presents particular difficulties from the well documented limitations that a human faces when observing and monitoring a collection of distributed concurrent processes Garcia Molina Germano amp Kohler 1989 Yet the need for tools supporting the deployment of complex systems is clearly recognised Rover Waheed Mutka amp Bakic 1998 propose a classification of the most typical tools categorised in debugging testing verification tools etc 2 1 Debugging Tools in Concurrent Parallel Programming Debugging a MAS is known to be a complex activity aimed at identifying structural and func tional errors Ndumu et al 1999 structural errors concern the organisation structure and relationships between agents while functional errors involves mistakes in the logic of the tasks to be performed There is however one third class of errors the coordination errors that occur when the individual agent behaviour is structurally and functionally correct but the emergent behaviour of the overall system is incorrect this is precisely where a suitable coordination infrastructure providing ad hoc tools can be of help Typically traditional debugging tools in concurrent and distributed programming provide features such as breakpoints on the single process traps to catch process exceptions single step tracing and explicit read write of
44. the interested reader to Denti amp Omicini 19992 3 Technology and Tools The tuple centre model is the core of the TuCSoN and LuCe coordination infrastructures for multi agent systems Omicini amp Zambonelli 1999 Denti amp Omicini 1999b both are based on Java as the standard glue technology are built upon the tuple centre notion and focus on the interaction aspects of a MAS so they are orthogonal and therefore unbounded to any specific agent model or platform Where the two infrastructures differ is in the abstraction of the network topology TuCSoN promotes network awareness providing an articulated topology model designed to support security and mobility issues Omicini amp Zambonelli 1999 Cremonini Omicini amp Zambonelli 1999 whereas LuCe is network transparent and specially suited for in tranet MAS Denti amp Omicini 1999b So though in the following we refer mainly to TuCSoN analogous considerations hold for LuCe too 3 1 Installation The adoption of the Java 2 platform as the implementation technology makes the installation of the TuCSoN LuCe infrastructures trivial all is needed is to download the tucson jar file to some directory of your choice Then a TuCSoN node can be activated Figure 1 By default the local node starts on the TuCSoN standard port 20504 but a different port can be specified on the command line the TuCSoN services can then be accessed via either the provided IDE tools
45. their full name of the form TupleCentreName TupleCentreIPAddress Here are some examples temperature 137 204 192 175 out data jan 15 17 5 14 default bongo unibo it out p 1 folder 5 lia org unibo it inp guests List Now suppose we want to change the default behaviour of a tuple centre re programming it so that each inserted tuple is automatically backed up transparently to user agents This can be done by adding via set_spec a suitable set of reaction tuples that perform the required backup operation like the following Figure 3 top temperature set_spec reaction out X out_r backup X If the specification is accepted you should get back the specification itself as an answer So all is needed to test the new behaviour is to perform an out operation for instance temperature out array index 1 value tomato To check whether the tuple centre behaved as expected the agent needs just to read a backup tuple temperature rd backup X hopefully getting backup array index 1 value tomato as answer Figure 3 bottom 3 2 2 The Inspector Tool The Inspector is a unique TuCSoN tool that enables the observation inspection and debug ging of the relevant information and activities related to the tuple centre communication and coordination states The key point here is that interaction between the Inspector and the observed tuple centre occurs at a different level the infrastructu
46. ult Task Result 6 Result lt perform Task 7 out result Task Result Table 1 Agent behaviour in the CNP Manager left and Contractor right 1 reaction out announcement Task out_r contract_state Task announced 2 reaction in bids Task _ pre in_r announcement Task out_r collect_bids Task reaction in bids Task L post in_r contract_state Task announced out_r contract_state Task evaluating L reaction out_r collect_bids Task L in_r collect_bids Task L in_r bid Task Bid out_r collect_bids Task Bid L reaction out_r collect_bids Task L in_r collect_bids Task L no_r bid Task _ out_r bids Task L 3 reaction out award Task Bid in_r contract_state Task evaluating L out_r contract_state Task awarded Bid out_r provide_responses Task L reaction out award Task Bid in_r award Task Bid reaction out_r provide_responses Task Bid _ rd_r contract_state Task awarded Bid out_r bid_response Task Bid awarded reaction out_r provide_responses Task Bid _ no_r contract_state Task awarded Bid out_r bid_response Task Bid not_awarded reaction out_r provide_responses Task _ List out_r provide_responses Task List reaction out_r provide_responses T L in_r provide_responses T L 4 reaction in bid_response Task _ awarded post in_r contract_state Task _
47. uple template Figure 5 bottom left as well as to log and save the possibly filtered tuple set shown Figure 5 bottom right so as to track the evolution as a sequence of tuple set snapshots Analogously the pending query view Figure 6 top left shows the set W of the pending in and rd queries along with the related information such as the requesting agent and the tuple centre local time when the request was received Filtering and logging of the set content work as above The specification view Figure 6 right shows the set S of the current reactions the tool includes an editor with syntax highlighting and syntax checking button check It is also possible to load a set of pre defined reactions and conversely to save the current reaction set to a file Once the desired reactions have been typed in or loaded the accept button sets them as the currently valid specifications replacing any previous set Finally the triggered reactions view shows the set Z of the currently triggered reactions This ability is crucial to debug coordination activities since tracking successful and failed reactions simplifies the understanding of the dynamics of the coordination flow inside the tuple centre For the sake of concreteness Figure 6 bottom left refers to the reaction specifications shown aside and reported here below in the case that an agent first outputs a p 4 tuple and then another agents emits a q 5 tuple reaction out
48. ve case coordination comes from the attitudes of each individual toward the organisations societies it belongs to while objective coordination prescinds from the subjective view of the coordinated agents and promotes the separation between the individual perception of coordination and the global coordination issues As a result objective coordination enables the modelling and shaping of the interaction space somehow independently of the interacting entities However just providing a suitably expressive objective or subjective coordination model is not enough to effectively support the engineering of Internet based systems suitable coordi nation technologies and run time support should be made available as well An efficient design development and deployment process requires suitable infrastructures providing designers and developers with a reliable support for commonly required application services for the above reasons coordination services should be regarded as fundamental ones To work with the widest range of hardware devices and software platforms including PDAs and embedded systems and to support the widest set of applications as well a coordination infrastructure for agent based applications should be easily deployable light weight and both statically and dynamically con figurable Also it should be Internet compliant and come along with effective methodologies and tools supporting the design development and debugging monitorin
49. web page reaction rd web_page Name Content pre in_r page hit_counter Name N N1 is N 1 out_r page_hit_counter Name N1 This kind of flexibility can be very useful for infrastructure maintenance since it simplifies both the monitoring and the detection of interaction related problems Coordination laws can also be exploited to prevent interactions that are invalid undesired in the current interaction context For instance a tuple representing some vital information such as the users of a system can be protected from unwanted removal i e from in operations by a reaction like the following reaction in users UserList post out_r users UserList In the same way insertion of critical information can be restricted to trusted sources reaction out price Who Info no_r source_authorised Who in_r price Who Info Although coordination laws can be exploited for security purposes security issues like the above should better be faced at a different meta level than the coordination level Cremonini et al 1999 provide more details on this 4 Case Study Here we briefly illustrate the impact and the benefits of the Inspector tool for the development and debugging of a simple version of a well known agent interaction protocol the Contract Net Protocol CNP Smith 1979 The CNP is one of the most common protocols for agent coordination Huhns amp Stephens 1999 generally used to solve the so calle
50. xperimented the benefits of the separation between agent computation and coordination Interaction specifications and protocols could be tested and debugged independently of agents either using special debugging agents or simu lating them via suitable tools as our Inspector In particular the role of tools seems crucial 3the out_tc DestTC Tuple operation which inserts a tuple Tuple into a tuple centre DestTC has been recently added to the ReSpecT primitives 16 in helping engineers harnessing the intrinsic complexity of agent interaction by providing them with the most effective views on the state and evolution over time of interaction within a MAS 6 Conclusions and Future Work Several proposals in literature face the issue of suitable debugging and testing tools for the deployment of agent based systems yet most of them seem still focused on the notions of process and agents trying to extend the tools found in sequential contexts toward interaction rather than focusing on the interaction dimension per se Recent works however outline the need of suitable infrastructure based tools that can provide support for inspecting monitoring and debugging interaction Graham McHugh Mersic Foster Windly Cleaver amp Decker 2001 In this work we identified the key aspects of such tools and discuss their development in the context of tuple based coordination infrastructures The TuCSoN infrastructure based on the ReSpecT model
51. ype get any new observation get only when update requested f update Figure 9 The CLIAgent of the manager announcing a new task and the related effect on the contract tuple centre triggered reactions and tuple set view 20 TuCSoN Command Line Interpreter command contracts s206 ingce unibo it rd announcement Task output waiting for tuple announcement Task answer announcement purchase book isbn 88 06 15304 8 TuCSoN Command Line Interpreter command contracts s206 ingce unibo it out bid purchase book isbn 88 06 15304 8 price 45 output waiting for tuple announcement Task answer announcement purchase book isbn 88 06 15304 8 ok TuCSoN Command Line Interpreter rcommand contracts s206 ingce unibo it in bid_response purchase_book isbn 88 06 15304 8 price 45 Answer output waiting for tuple announcement Task answer announcement purchase_book isbn 88 06 15304 8 jok waiting for tuple bid response purchase book isbn 88 06 15304 8 price 45 Answer Figure 10 The CLIAgent of a bidder after executing some interaction steps reading the an nouncement providing a bid and waiting for the answer 21 1024346392054 1024346319098 bid purchase book isbn 88 06 15304 8 price 50 bid purchase _book isbn 88 06 15304 8 price 45 bid purchas

Coordination Tools for MAS Development and Deployment

Contents

Download Pdf Manuals

Related Search

Related Contents