The CAST Manual - Intelligent Agents Laboratory
Contents
1. output link dtank move status complete turn direction clockwise counter clockwise status complete 2 Build domain adaptor According to the interaction protocol you need build a domain adaptor that specifies basic communication with the domain In the dTank example the adaptor is composed of four classes e cast3 domain JTankWorld TankClient defines the socket connection functions that facilitate the basic communication between the world simulator and agents public TankClient String host int port try this host host this port port tankSocket new Socket host port out new PrintWriter tankSocket getOutputStream true in new BufferedReader new InputStreamReader tankSocket getInputStream 29 37 catch UnknownHostException e System err printIn Couldn t connect to hostlport host I port System exit 1 catch IOException e System err printIn Couldn t get I O for the connection System exit 1 send agent join this start e cast3 domain JTankWorld AgentJTankWorldAdapter defines how to apply the operators such as move and turn and how to handle the incoming messages public void doMove operatorID set command base dTank_delim tail dTank_delim id sendWorldMessage moveForwardl operatorID public void doTurn String dir operatorID int direction 1 if dir equalsIgnoreCase left2. 2x y 2z e mod x y z e pow x y z Here are some examples of using these math predicates gt query 1 2 3 1 2 3 2 3 x 5 0 3 x 4 0 4 gt query mod 9 4 x mod 9 4 1 0 Recently we added some other useful math predicates like sqrt exp log mod sin cos and tan Another class of procedural attachments are provided for support of list operations In Jare lists can be represented as terms within predicates by using a special function list that can have any length For example the list of a b and c is list a b c and the empty list is just list Currently procedural attachments are provided only for basic operations such as cons first and rest These allow you to construct and deconstruct lists Then a larger set of additional list predicates is defined over this through axioms in list kb which is automatically loaded into Jare on startup Cons works as follows You must give it an object and a list and it returns a new list with the object stuck in the front Hence cons takes a total of 3 arguments If the 3rd argument is bound cons verifies its correctness i e succeeds if it is correct else fails More interestingly if the 3rd arg is unbound Jare binds the answer new list to it First and Rest extract the corresponding elements from a list returning an object or a list respectively Again if the second argument is bound its correctness is ch
3. CONSTRAINTS lt Cond gt The FORALL construct is an implied PAR over the condition bindings whereas the FOREACH is an implied SEQ over the condition bindings These process types are fairly expressive when the number of choices is unknown before runtime FOREACH is most useful in giving a single agent a list of tasks with different arguments and FORALL can be used when multiple agents are needed but their numbers are unknown The CHOICE construct is a control structure that takes a list of processes and executes them in order until one completes successfully The AGENTBIND construct introduces flexibility to the teamwork process in that the agent selection is made dynamically at runtime to assure the satisfaction of certain teamwork constraints such as finding an agent that is capable of some operation The values for agent variables are to be assigned so as to satisfy the constraints For example AGENTBIND f1 CONSTRAINTS playsRole fi fighter closestToEnemy f1 states that fi will be assigned to an agent that plays the role of fighter and is closest to the enemy The selected agents are then responsible for performing later steps operators sub plans or processes associated with the agent variables The scope for the binding to an agent variable extends to either the end of the plan in which the variable appears or the beginning of the next agent bind statement that binds the same variable whichever comes first An AG
4. plays role Fl fighter_tank plays role F2 fighter_tank role fighter_tank move turn attack Moreover the agent should also specify how the tasks should be allocated among the team members plan attack_enemy pre cond at tank x y is enemy tank process agent bind 00 constraints playsRole fighter_tank f endproc endplan 5 Starters A CAST team is normally configured in an XML file which defines the agent name domain MALLET plan cast monitor knowledge base type world address port and communication type lt xml version 1 0 encoding UTF 8 gt lt DOCTYPE CAST SYSTEM jTank dtd gt lt CAST gt lt AGENT gt lt NAME gt F1 lt NAME gt lt DOMAIN gt cast3 domain JTank World JTankDomain lt DOMAIN gt lt MALLET gt teamplans jTank casttankteam mlt lt MALLET gt lt MONITOR gt localhost lt MONITOR gt lt KBTYPE gt cast3 util Jare JareKBWrapper lt KBTYPE gt lt WORLD_HOST gt localhost lt WORLD_HOST gt lt WORLD_HOST_PORT gt 3400 lt WORLD_HOST_PORT gt lt COMM_FEATURE gt PROTELL lt COMM_FEATURE gt 33 37 lt AGENT gt lt AGENT gt lt NAME gt F2 lt NAME gt lt DOMAIN gt cast3 domain JTank World JTankDomain lt DOMAIN gt lt MALLET gt teamplans jTank casttankteam mlt lt MALLET gt lt MONITOR gt localhost lt MONITOR gt lt KBTYPE gt cast3 util Jare JareKBWrapper lt KBTYPE gt lt WORLD_HOST gt localhost lt WORLD_HOST gt lt WORLD_HOST_PORT gt 3400
5. AgentOrTeamName Cond lt START gt AgentOrTeamName Invocation lt CAPABILITY gt AgentName AgentName Invocation Invocation my my lt ROLE gt RoleName Invocation Invocation lt PLAYSROLE gt AgentName RoleName RoleName lt FULFILLEDB Y gt RoleName AgentName AgentName 35 37 m PreConditionList lt PRECOND gt Cond IF FALSE if false lt FAIL gt lt WAIT gt lt IDENTIFIER gt lt ACHIEVE gt EffectsList lt EFFECTS gt Cond TermConditionsList lt TERMCOND gt lt SUCCESS gt lt FAILURE gt Cond NumSpec lt NUM gt lt EQ gt lt LT gt lt GT gt lt LE gt lt GE gt lt IDENTIFIER gt n IOperDef lt IOPER gt OperName VariableListOpt PreConditionList EffectsList Or TOperDef lt TOPER gt OperName VariableListOpt PreConditionList EffectsList NumSpec PlanDef lt PLAN gt PlanName VariableListOpt PreConditionList EffectsList TermConditionsList lt PROCESS gt MalletProcess ByWhomSpec AgentOrTeamName MixedListReq lt VARIABLE gt MalletProcess lt SEQ gt MalletProcess lt PAR gt MalletProcess lt IF gt lt COND gt Cond M
6. Start new plan Figure 3 Individual agent monitor 6 2 1 KB query panel The KB query panel allows a user to access agent s knowledge base during runtime Therefore it provides a powerful tool for agent developers to debug watch and modify agent s knowledge agent behavior dynamically Figure 4 shows a sample out put and the list of query functions which is explained below the Figure 24 37 Agent Name F1 Simtabon Domain Cust domuin rostuuworid HescueDo mun Show MALLET Specifications Save as MALLET Fite cust util Jare JaroKBWrapper capabilites icapabiity F 7 move ash ersnquish Interactions Recower iD fi Message Type contro Message Text Start new plan Figure 4 KB query panel Query is for querying knowledge from the knowledge base Input the sentence in the blank space with two brackets and then press Enter The output of the query will be displayed in the output window If it is true the queried sentence will be displayed again If it is false fail will be displayed Assert 1s for inserting a new knowledge to the knowledge base Input the sentence in the blank space with a bracket and then press Enter Asserting S will be displayed in the output window where S means the asserted sentence QueryforBindings is for querying for agent bindings Retract is for deleting a fact from the knowledge base It is a reverse pro
7. direction 0 sendWorldMessage rotatel direction a operatorID public void handleMessage String message String host int port super handleMessage message host port handleWorldMessage message public void handleWorldMessage String msg if agent null agent handleWorldMessage msg public void sendWorldMessage String msg send msg e cast3 domain JTankWorld JTankAgent defines how to handle parse the incoming messages public void handle WorldMessage Object message String strMeg trimBye String message StringTokenizer strtok new StringTokenizer strMeg stop 30 37 String type new String if strtok hasMoreTokens type strtok nextToken else System out printIn error getting world message message toString if type equalsIgnoreCase InitialSettings handleInitialMessage strtok else if type equalsIgnoreCase ACTION handleActionMessage strtok else if type equalsIgnoreCase ACK handleAckMessage strtok else if type equalsIgnoreCase STATUS handleStatusMessage strtok else if type equalsIgnoreCase INPUT handleInputMessage strtok else if type equalsIgnoreCase SCAN handleScanMessage strtok else if type equalsIgnoreCase EVENT handleEventMessage strtok else if type equalsIgnoreCase VISUAL handle VisualMessage strtok else if type equalsIgnoreCase STOPAGENTI setPaused true else if type
8. which are captured by its information flow table e Upon acquiring new information sensed from environment or derived from effects 1 Based on the common knowledge of observabilities expressed in JARE its information flow table and other beliefs an agent determines whether the 12 37 information needers among its teammates can get the same information by themselves and hence the agent doesn t need to inform them When considering communication cost risk an agent will decide whether it should proactively inform the new acquired information to the needers by evaluating the utility of protell vs not protell e When lacking of certain information 1 An agent can determine the potential information providers by checking its information flow table When considering communication cost risk an agent will decide whether it should proactively ask an information provider by evaluating the utility of ask vs not ask e When executing along the shared PrT nets 1 4 By analyzing the currently active PrT nets determine the actions within the team that are to be performed next refer to algorithm 1 In the case of team operations or joint AND XOR OR operations synchronize with teammates as necessary Upon dynamic agent assignment determine which agents will be dynamically assigned to some actions lacking pre specified performers this may involve coordinating with other agents Determine whether o
9. file This reduces the amount of setup needed as a user can simply use the XML file to specify what agents will be started and their attributes along with a few CAST system attributes To Run CAST agents If you have installed ant Type ant jtankstart under the sub folder Cast If you do not have ant installed Type run_jtank or Type java classpath lib lib cast jar classes Djava security policy cast policy cast3 monitor CastMonitor config jTank jTank xml cast3 domain JTankWorld JTankA gentStarter under the sub folder run Testing After you run the CAST agents in the agent monitor you may click the button unpause all to run them Configuration 4 37 Under run folder there are three sub folders config images and teamplans Config folder consists of the configuration files of the team which are saved in XML files Images folder consists of the images for displaying the world simulator Teamplans folder consists of the plans of the team which are saved as MALLET files Cast agents are configured with an XML file e g config jTank jTank xml which contains the configuration of the team of the agents For every agent we should explicitly indicate at least the NAME DOMAIN MALLET MONITOR KBTYPE WORLD_HOST and WORLD_HOST_PORT MALLET Plan In the example tank scenario the procedural knowledge contains a two phase plan search and attack Initially a team wanders around and searches for ene
10. lt WORLD_HOST_PORT gt lt COMM_FEATURE gt PROTELL lt COMM_FEATURE gt lt AGENT gt lt CAST gt An agent starter class cast3 domain JTankWorld JTankA gentStarter is responsible for instantiating the agents while it hasNext Element agent Element it next String name agent getChildText NAME String domain agent getChildText DOMAIN String mallet agent getChildText MALLET String monitor agent getChildText MONITOR domainhost agent getChildText WORLD_HOST domainhostport agent getChildText WORLD_HOST_PORT String feature agent getChildText COMM_FEATURE JTankAgent cast new JTankAgent name mallet domainhost domainhostport cast setFeature feature AgentDisplay ap new AgentDisplay cast cast3 util Jare JareK B Wrapper manager addA gent name ap 6 Test Observing the agent s behavior checking agent s knowledge base with query commands and monitoring the state of the agent s process net are basic methods for testing and debugging The followings are general guidelines for testing e Test the agent adaptor before testing the agent e Test the declarative knowledge predicates offline before testing the plans e Test plans with simple conditions before testing with complex conditions e Test single agent before testing the team 34 37 8 Appendix A the syntax of MALLET CompilationUnit PlanName OperName PlanOrOperName AgentName T
11. solutions week day x day x not weekend x The next major feature of Jare is procedural attachments Procedural attachments are predicates with special pre defined meanings that are essentially implemented in Jare code Currently there are two classes of procedural attachments math operations and list operations Math operations include two types relations e g equality and functions e g Relations are generally 2 argument predicates Both arguments must be bound i e you can t evaluate a relation on a free variable Jare will signal an error if you try and both arguments must be numeric Here is a list of relational predicates 2x 2y e lt x y e gt 2x y e lt x y e gt 2x y Straight inequality may be implemented as not x y Note Jare attempts to convert anything that begins with a or digit into a float and math operations are carried out on these so 6 0 and 6 are equal However and may still be used as symbols themselves The predicate eq can be used for comparing equality of symbols 20 37 Math functions generally take 3 arguments They can be used to verify relations if all 3 args are bound or any one of the three may be left unbound and Jare will fill in the answer expect mod and pow for which only the third argument can be unbound The implemented math functions are e 2x y 22 e 2x y 2z e 2x y 2z e
12. specification or indirectly through a role specification For example to state that FO and UO have the same capabilities and are able to perform actions move and detectEnemy one would have capability FO UO move dir detectEnemy Alternatively one could state role fighter move dir detectEnemy shoot x y role bomber move dir detectEnemy co_fire x y where fighters have the additional capability to shoot an enemy fighter and bombers have an additional capability to bomb an enemy base through the co_fire team operator 4 2 Process knowledge Operators are the basis for the hierarchical construction of team plans i e operators are atomic actions in a plan hierarchy Plans essentially characterize team processes A plan in MALLET is composed of precondition effects and termination conditions constraints for task allocations and the procedural description of the process The precondition represented by a logical conjunct expresses a necessary condition under which a plan or an operator can be performed The effects also represented by a logical conjunct state the conditions that hold after the plan is successfully accomplished The termination conditions describe two kinds of conditions for terminating a plan the success conditions and the failure conditions The constraints for task allocation are specified by the AGENTBIND construct The process of a plan describes the procedure of how a team wil
13. the next step of the team plan The parallel process in MALLET currently specifies parallel activities in a team but not parallel activities for an agent Therefore each agent can choose only one parallel branch to enter If an agent can enter multiple parallel branches the current implementation of CAST randomly chooses one of these branches This selection scheme is simple but does not address the issue of potential deadlocks For instance suppose three agents A B and C enter a parallel process with three branches Suppose the first branch invokes a team plan that requires A and B the second branch requires B and C the third branch requires C Assuming agent A chooses the first branch agent B chooses the second branch and agent C chooses the third branch Only agent C will be able to proceed while agent A and B each waits for another teammate indefinitely to execute their branch 11 37 This problem can be addressed in the future in two ways 1 by developing algorithms to detect potential deadlocks in a MALLET team plan specification and 2 by improving the coordination among agents in choosing parallel branches The CAST kernel includes a set of algorithms that CAST agents can use to determine the actions including domain actions and communication actions to be performed at each time step Such algorithms include PrT net interpreter DAA dynamic agent assignment DIARG protell based on information flow table decision the
14. 127 0 0 1 1120 127 0 0 1 1122 130 203 156 37 1725 kuan 203 LAR Figure 7 dTank plan view monitoring interface Domain monitor Domain monitor is normally provided for human to monitor the online domain situation Therefore the information coming to the domain monitor may not be appropriate to feed as the agent s sensory input For example one of the goals of dTank is to provide a testing environment in which agents and humans can interact on equal footing both synthetic and human players should have similar sensory input from the game environment If agents are able to view the entire game board along with the positions of all tanks while human can only see other tanks within their visual field agents would have an unfair advantage over their human opponents CAST monitor CAST monitor is useful to monitor the state of agents It also provides log functions that can keep track of the agents operational behavior However CAST monitor is not essential for applying the CAST agents in a domain CAST source code MALLET 2 includes the code for approaching MALLET files Domain is the world simulator The operators in the MALLET file should have been defined in domain Kernel2 the key part of CAST agent which realized the main features of CAST such as Dynamic Task Allocation Detection of Information Needs of Teammates Proactively Information Delivery Monitor is to display the interface to the user for dynamically contr
15. ENTBIND statement of an agent variable can occur anywhere in a plan as long as it precedes the first DO statement in which the agent variable is an actor The DO statement in a MALLET plan provides two functions 1 it invokes a process 2 it specifies the actor s for the process Syntactically the actor specification can be a list of agent variables or the name of a team subteam Semantically the agents included in the actor specification should be sufficient for accomplishing the process From teamwork perspective the actor specification in MALLET enables a knowledge engineer to describe the subteam for carrying out a team plan involving multiple agents The JOINTDO construct provides a means for describing multiple synchronous processes to be performed by the identified agents or teams in accordance with the specified share type A share type is either AND OR or XOR For an AND share type all of the specified subprocesses must be executed For an XOR exactly one subprocess must be executed and for an OR one or more subprocesses must be executed A JOINTDO statement is not executed until all involved team members have reached this point in their plans Furthermore the statement following a JOJNTDO statement in the team process does not begin until all involved team members have completed their part of the JOINTDO 17 37 5 JARE JARE Java Automated Reasoning Engine is a back chaining theorem prover for making inferences usi
16. RE QUE ty panel aiii 24 6 2 2 Agent communication panel iii denied 25 6 2 3 PrE Met display panel es edu a Belge ete io 2 MOTO POOL RES 27 TI CAST System OVervieW usd id 27 As AA svar secetacoveaann 27 ES 28 CAST MOM isis 28 CAST source code A tee Dee eee sabes 28 7 2 Basic steps of Building a CAST team oy sciscesvecesalessavsseesddssctns darias dress inician 29 1 Study domain and get rules APIS EC ii a BAG 29 Ze B ld domami AGA DIOR ss isie aee a cavqechdsdeedtla aemean edad ees 29 3 Get domain knowledge A iengsuas DeGecaandes dedeesteaeacse tut eeiaeasucvaasess 32 4 Teamwork knowledge eeeesseessesseeeseeeeseeessesseesseesseeesseeesseesseessersseeessseesseest 33 PE T E A E A E Ihde ek re 33 6n TEST peso Getta a E E A E CE E EE Oe E a 34 8 Appendix A the syntax of MALLET ssessssseesseessesssesssesessseessresseesserssseessseessesseessee 35 9 References Geia a a a a e a e ote E ee de ott 37 2 37 1 Preface CAST Collaborative Agents for Simulating Teamwork is a teamwork agent architecture that enables building collaborative agents as teams This manual gives a high level introduction on CAST architecture how to use CAST and how to program with CAST Readers may refer to 1 for more information on it from a theoretical perspective John Yen is the main architect for CAST CAST version 1 0 was developed by Jianwen Yin Michael S Miller has developed the CAST version 2 0 that is currently us
17. ST kernel The kernel of CAST architecture includes a set of integrated algorithms goal selection dynamic task allocation communication decision It also controls the agent sensing decision making and action operation processes Dynamic agent assignment During the execution process agents select tasks to execute according to the PrT net and the beliefs about the team and domain If the selected task is a sub plan the agent will set it as the current task that the team needs to accomplish At the same time the agent will try to coordinate with other agents who are also assigned to this task The plan is represented in a PrT net that is generated from MALLET The algorithm for task allocation in CAST is called dynamic agent assignment DAA DAA first determines the current goal and finds an associated plan using the goal selection process Each plan defined in MALLET specifies the constraint conditions on how the tasks in this plan are allocated Such conditions may include role constraints workload conditions or domain related conditions Before executing a task each agent 8 37 will check these constraints with its current beliefs about the domain or teammates and determine the suitable agents whose conditions satisfy the constraints Agents workload status will also be considered during this process The result of the DAA will determine which agents are going to do the task and will play an important role in updating beli
18. The CAST Manual John Yen Thomas R loerger Michael S Miller Shuang Sun Kaivan Kamali Shizhuo Zhu Xiaocong Fan Richard A Volz Technical Report No XXXXX School of Information Sciences and Technology The Pennsylvania State University May 11 2004 324 IST Building University Park PA 16802 814 865 6178 Tel 814 865 5604 Fax Department of Computer Science Texas A amp M University College Station TX 77843 Table of contents A ech ccd cae A A EA EAE aAA AETA EENT cae adn hs 3 2 Getting stats 4 3 About CAS vi ia 6 SOC a e ds 6 A a O 6 3 3 CAST architecture eeren E es 6 Teamwork knowledge representation esscceessecessceceenceceenceceenceceeceecseeeecseeeeeaeees 7 VARESE Da ca dade a Daas e ea da 8 CAST kernel a n a a A ER E e aid 8 Dynamic agent asii 8 Communication decision s scsessecasescatsnsvcvcatasevsadeassnenede reacios niente cada 9 o A E ogc uses 9 Agent COMU CO A AAA AI AA E Bleed 10 Agent interaction With Comain cecceessccesssecessecesaeceeseeceeaeeccsseeecseeeecseeeeeeteeeesaes 10 EX le le dd e a ES 11 AMABA A ia Os 14 4 1 cie A nda cova cas cascada va dade a a a A a aaie 14 4 2 Process KNOW liado is 15 Bie DARE os eiess cea supeas ca leaiuacaseavas oes iets dd 18 LASA of JARE langage ii 18 S 2A ist Of TEATES O 20 6 How to use CAST sita i E anaa AEAEE lotes riia i basae 23 6 1 CAST team DOMO 23 6 2 GIVI a al AGE ME MONIT ni EE A gee 24 Gl
19. ain adapter The teamwork knowledge specification allows agent designers to use MALLET to design team structures and team process of the team At the same time domain knowledge is captured in the agent s world model The coordinated plan execution module makes action decisions and communication decisions The CAST agents operate in a distributed fashion and they use JAVA Remote Method Invocation RMI for communication Each 6 37 agent interacts with domain through a domain adapter which integrates the four other components and makes CAST agents adaptive to different domains An overview of the CAST agent architecture is shown in Figure 1 ae ae am a0 SP SEEPS SEE OOO AROS II aa TO TO wesw ere essere see wee mea ece I Coordinated Team Plan Execution Action Execution Communication Domain Action Domain Simulator Domain World lt Bomain Adapter l l Fact Base amp z i F q Agent y i Agents ie gt Domain Scenario Generator Communication Server lt Domain Adapter gt Figure 1 The CAST agent architecture Teamwork knowledge representation Agent designers can use the team specification language MALLET to define team structures and team plans MALLET supports most of the crucial elements of defining a team s organizational structure such as members of the team roles and responsibilities capabilities of agents sub teams etc Team structure information once loaded i
20. alletProcess MalletProcess lt WHILE gt lt COND gt Cond MalletProcess lt FOREACH gt lt COND gt Cond MalletProcess lt FORALL gt lt COND gt Cond MalletProcess lt CHOICE gt MalletProcess lt JOINTDO gt lt AND gt lt OR gt lt XOR gt ByWhomSpec MalletProcess Hye n PlanOrOperName lt IDENTIFIER gt lt VARIABLE gt lt DO gt ByWhomSpec MalletProcess lt AGENTBIND gt VariableListReq lt CONSTRAINTS gt Cond my AssertDef RetractDef 36 37 9 References 1 2 3 4 5 6 J Yen X Fan S Sun R Wang C Chen K Kamali M Milller and R A Volz On Modeling and Simulating Agent Teamwork in CAST presented at 2nd International Conference on Active Media Technology Chongqing P R China 2003 D Xu R A Volz T R loerger and J Yen Modeling and Analyzing Multi Agent Behaviors Using Predicate Transition Nets International Journal of Software Engineering and Knowledge Engineering vol 13 pp 103 124 2003 J Yen X Fan and R A Volz On Proactive Delivery of Needed Information to Teammates presented at AAMAS 2002 Workshop of Teamwork and Coalition Formation Italy 2002 J Yen X Fan and R A Volz Proactive Information Exchanges Based on the Awareness of Teammates Information Nee
21. as been used to study how to enhance team performance by proactive information delivery The experiments showed agents coordinate with different mental models on teamwork have different team performance Second CAST agents can be used as intelligent agents to model opponents in a computer game For example CAST has been used in dTank a tank game environment to compare the team performance of different agent architecture Third CAST agents can be used to simulator team members in a training environment CAST has been used in a simulated combat environment that designed for combat operational training In the training CAST agents are used to provide online feedbacks to trainees Of course CAST can be applied to other problem domains where teamwork or collaboration is required In this section we use dTank domain as an example to illustrate how to build an agent team with CAST Figure 6 shows the system architecture including domain domain monitor agents and agent monitor From the Figure we can see it is a distributed system which is composed of two parts domain and agents Agents interact with a domain sense domain states and apply operation effects to the domain Agents interact with each other by communication A CAST monitor provides a display for users about the agents knowledge states and operations Normally a domain monitor is linked to the domain or simulator to show the dynamics of domain states Figure 6 CAST system archite
22. cess of Assert Execute is for dynamically loading a file New info is for triggering a pro tell prefixed as info Print KB is for printing out the knowledge base in the output window 6 2 2 Agent communication panel The part at the bottom of the window provides interactions between different agents Receiver ID identifies the receiver Message Type identifies the type of a message you want to deliver to other agent Message Text is the content of a message Start new plan identifies the other plan you want to start at this point 6 2 3 PrT net display panel When the Show Process Nets button on the left side is clicked all the plans of this agent will be shown in the form of the Process Net The process net of a plan is a PrT net as Figure 4 shows A PrT net consists of Places Transitions and Arcs A place in oval indicates a possible system state A place can contain tokens to indicate current state of 25 37 the system A transition in rectangle is an action which changes the system state from one place to another A transition is connected to a place by an arc The arcs starting at a place and ending at a transition are called input arcs and those starting at a transition and ending at a place are called output arcs The display of a Pr T net uses five different colors to indicate different states e Grey the grey rectangle separated with the process net at the top of the window indicates th
23. cture Domain Domain simulator dTank acs ist psu edu dTank is a simulator created by the Applied Cognitive Science Lab ACS at the Pennsylvania State University It provides an adversarial real time test bed for conducting experiments with cognitive models in a tank combat setting The simulator constructs a 10 by 10 grid world composed of stone barriers and grassland A simulated tank can move forward on grassland but cannot cross a stone or another tank Tanks are able to navigate their environment by moving in the forward direction or turning either clockwise or counter clockwise A tank can rotate its turret freely 3600 and the visual field is centered on the turret angle It can observe world objects other tanks or stones if the objects are within its 1000 visual field without being blocked by other game objects A tank aims its gun at an enemy tank by turning its turret Initially a tank has a certain level of health represented by health points which decrees after the 21187 tank is shot by others When health points drop to zero the tank dies A tank can communicate with its teammates through a simulated radio function A first person driving view that can log behavior is also included which will support human vs human and human vs agent games as well as o data to model human users Figure 7 shows a plan view interface for monitoring a tank game The dTank manual 6 provides more information dank fala ES
24. d participate in a team action know that the operation has been completed and it is permissible to proceed A future optimization could take observability into account and only inform teammates who were believed unable to observe the performance of the action The actual agent communication is placed in a module outside the CAST kernel because the lowest level mechanisms used are domain dependent In our test simulation worlds we have used JAVA RMI and socket level communications In a domain in which CAST is integrated into an operational setting e g to provide assistance to live human operators other forms of communications e g voice images and video might come into play as well For purposes of this paper we simply assume that the communication module provides reliable communication among team members when called upon to do so Most importantly on the receiving end it sends received information to a receiving agent within the Kernel that can interpret the information received and update the JARE knowledge base Agent interaction with domain The domain or simulation environment is independent of CAST kernel A typical domain contains its fact base and some policies To interact with the domain a CAST agent requires that domain environment periodically provide new sensing information and accept agent domain action commands Each agent has its own instance of the CAST Kernel and domain adapter and interacts with the domain ind
25. ds presented at AAMAS 2003 Workshop on Agent Communication Languages and Communication Policies Melbourne Australia 2003 T Finin R Eritzson D McKay and R McEntire KQML as an agent communication language presented at Third International Conference on Information and Knowledge Management CIKM 94 1994 I G Councill G P Morgan and F E Ritter dTank A Competitive Enviorment for Distributed Agents The Pennsylvania State University University Park Technical Report ACS2004 1 03 30 2004 2004 37 37
26. e agents can anticipate what information other agents might need DIARG is built on top of the theory for proactive information exchanges 4 and is used to anticipate action performing information needs The on line component of DIARG monitors information newly sensed by an agent to determine whether it matches information needs of teammates If there is a match the agent decides whether to proactively inform teammates Communications could introduce risks to the agent e g that an enemy could intercept the communication and learn information potentially harmful to the team Therefore agents should consider the risk as part of the communication cost in making their decisions on whether to proactively communicate with teammate or not Domain adapter The domain adapter acts as an interface to the simulation environment Once agents make all of the decisions they can carry out action operators The CAST agent kernel does not define domain operators A PrT net is executed at the abstracted operator level The execution of the operators will invoke a set of separate domain operator definition class which is integrated with a domain adapter The function of the domain adapter is to provide domain application layer communication such as API functions or messages The domain adapters are implemented by dynamic class loading and can consist of arbitrary Java code The domain definition class and domain adapter make CAST architecture adaptable to not on
27. e local binding of the variables of the agent states E g FIRE fire2 indicates the FIRE variable is bound to fire2 at the point e Red some places will be colorized as red when they are active e Green the places that are not active e Yellow the transitions in yellow means they have already been passed through before e Blue the transition in blue means they haven t been passed through yet The three buttons listed below the process net window are Fire a Transition Toggle Layout and Print Net Fire a Transition provides a convenient way to test how the process will be executed That means the agent developers can check if the agent can make a correct decision by comparing with the desired choices and the actual choices Toggle Layout button is used for automatically toggling the layout of the Pr T Print Net is for printing out the Pr T an son oro Process Not Plan eat Procase Not Plan Process Net m sect moveto plan roscueF ire Process Net Plan rencusOni im Process Net Plan workOnk ir a Proces Not Plan rmoveTaFke Progen Not Plan moveToFim Procesa Net Plan MeFighting Fire a Transition Toggle Layout Print Nat lt gt Figure 5 PrT net display 26 37 7 How to program in CAST 7 1 CAST system overview A team of CAST agents can interact with different types of domain or environments First CAST agents can be used as a model to simulate team behaviors For example CAST h
28. e process knowledge describes what the team is planning to do and a plan of how the team is to accomplish its goals All agents have a copy of these plans and hence know what is to be done Second a MALLET parser compiles the teamwork knowledge into a PrT Net representation which is an internal representation of the agent s SMM about the status of the team s process PrT nets as a high level formalism are more expressive than traditional PrT nets i e place transitions nets 2 The third component of the agent s SMM is a knowledge base that reasons about the agent s belief regarding the world and the structure of the team This knowledge base also contains domain knowledge of the agents The knowledge base of an agent is initialized by facts and domain knowledge known to the agent However it is continuously updated by sensor inputs and communication messages received by the agent Therefore even if a team of agents starts with the same knowledge base they will evolve into different but overlapping ones as they sense different information from the environment and receive different messages from teammates 3 3 CAST architecture The CAST framework is designed to model well structured agent teams and to be able to adapt to dynamic environments There are five major integrated components in the CAST architecture to support these objectives teamwork knowledge specification coordinated plan execution world model communication and dom
29. eamName AgentDef TeamDef MemberOf GoalDef Start CapabilityDef RoleDef PlaysRole FulfilledBy IOperDef TOperDef PlanDef RuleDecl LoadDecl lt EOF gt lt IDENTIFIER gt lt IDENTIFIER gt lt IDENTIFIER gt lt IDENTIFIER gt lt IDENTIFIER gt AgentOrTeamName lt IDENTIFIER gt RoleName IdentifierListReq VariableListOpt VariableListReq MixedListOpt MixedListReq Invocation AgentDef TeamDef MemberOf Pred Cond AssertDef RuleDecl LoadDecl RetractDef GoalDef Start Capability Def RoleDef PlaysRole FulfilledBy lt IDENTIFIER gt 2 lt IDENTIFIER gt 2 lt VARIABLE gt 2 lt VARIABLE gt lt IDENTIFIER gt lt VARIABLE gt 2 lt IDENTIFIER gt lt VARIABLE gt PlanOrOperName lt IDENTIFIER gt lt VARIABLE gt lt AGENT gt AgentName lt TEAM gt TeamName AgentName lt MEMBEROF gt AgentName TeamName TeamName lt IDENTIFIER gt lt NOT gt lt EQUATION gt lt LT gt lt GT gt lt LE gt lt GE gt lt IDENTIFIER gt lt VARIABLE gt Pred Pred lt NOT gt Cond lt ASSERT gt Pred lt RULE gt Pred lt LOAD gt lt IDENTIFIER gt lt RETRACT gt Pred lt GOAL gt
30. ecked else the variable is bound to the answer Here are the basic list predicates with examples e cons x y z cons x list a b list x a b e first x y first list x a b x e rest x y rest list x a b list a b 21 37 The additional list predicates defined through axioms include e empty x e length x n e nth x n y n is the position starts counting from 0 e second x y e third x y e member x y checks to see if x is in the list y e reverse x y e append x y z e g append list 1 2 list a b list 1 2 a b Generally only the last argument can be unbound for these predicates 22 37 6 How to use CAST CAST agents can be displayed and controlled through CAST monitor a graphical user interface The CAST monitor has two parts team monitor and individual agents monitor The team monitor is used for controlling the team actions viewing or adding team members starting or suspending team perorations and displaying the behavior of the team by listing the actions for each agent see Figure 2 Each agent has an individual agent monitor a tab captioned as the agent name In addition to the functions that included in the team monitor the individual agent monitor provides other comprehensive monitoring functions including KB query panel that allows a user to access agent s knowledge base a PrT net display that allows a user to monitor the current proc
31. ed Thomas R Ioerger has coded JARE At the Penn State University CAST has been future extended on the basis of version 2 0 and different domain scenarios have been implemented and tested with CAST agents 3 37 2 Getting started System requirements CAST v2 0 is developed with Java the only requirement is JDK1 3 1 or J2SE v 1 3 1 http java sun com j2se 1 3 However we recommend you install ant http ant apache org which will makes compiling and running easy Installation When the package is unzipped it will automatically create and copy the required files in the directory in which it was unzipped You need to start the dTank simulator an example domain under dTank 2 0 8a before running the test A dTank manual is available at http acs ist psu edu dTank Compiling Use the ANT build xml file in the cast3 src directory In this file there are a number of targets for both compiling and executing the CAST software If ANT is set up then typing ant projecthelp returns the following entries Buildfile build xml Main targets e Build Builds jar file e Cast all Runs CAST with example XML file e Clean Removes all classes and jar files e Docs Creates Java docs Default target build Running There are the targets that can be executed using ANT ANT can both compile and execute cast3 Typing ant with no arguments will compile the code CAST has an XML parser that it uses to parse a configuration
32. efs regarding information needs that agents use to provide information proactively for their teammates Communication decision The algorithm used to infer information needs is called the Dynamic Inter Agent Rule Generator DIARG DIARG is the most crucial part of CAST agent in terms of achieving the proactive assistance behavior in team environment The foundation of DIARG algorithm has been established by extending the SharedPlan theory with the formal semantics of proactive communicative actions 3 4 We have shown that an agent s consideration of proactive assist behaviors can be derived from axioms in the formal framework A team s SMM about the information needs of its members is captured in the framework by a model operator InfoNeed A J t Cn which represents that agent A needs information at time under the context Cn The arguments of these model operators correspond to attributes of the information need table which serves as a computational SMM about information needs for a team of agents DIARG has an off line component and an on line component The off line component generates the information need table using teamwork knowledge specified in MALLET The table contains information requirements preconditions for each task and candidate agents who are potentially responsible for the task When an agent is dynamically assigned to a task the content of the table will be updated according to the result of DAA Using this tabl
33. ependent of the interaction of other agents with the domain The CAST Kernel is designed to be domain independent so that CAST can be readily adapted to different domains In order to use CAST with a specific domain one needs to do three things 1 develop a domain specific adapter 2 encode necessary domain knowledge in JARE rules and 3 develop the teamwork plans which will usually have domain specific aspects CAST provides a template which can be used as a basis for 10 37 developing the domain specific adapters Domain operations that agents can perform are specified in an interface and the adapter must provide an implementation The operations used in MALLET programs must conform to the interface CAST then simply determines the operation names and parameter profile by parsing the MALLET plan and invokes the methods through the interface when it determines that an operation should be performed The interface also specifies methods by which domain information can be acquired to satisfy JARE queries Execution An important issue for CAST kernel to address is to achieve desired coordination and synchronization in a distributed manner Synchronization for a team operator is established using the following conversation policy 1 Each agent involved in the team operator i e an actor sends a synchronization message to other actors of the operator when it is ready to perform the operator 1 e after it evaluates the operator s p
34. equalsIgnoreCase STARTAGENTI setPaused false else if type startsWith join handleUndefinedMessage type strtok e cast3 domain JTankWorld JTankDomain defines how to process the sensed information and how to apply operators the operators will be called by the agent execution module public void processSense if setTeam jTankWorldAdapter sendWorldMessage setTeaml getTeam setTeam true updateTime KBEnv solution new KBEnv String find sense sense solution getAgent getKB queryForBindings find while solution null Vector sense Vector solution lookup sense 31 37 getAgent proactiveTell sense solution getAgent getKB nextQ JareKBWrapper getA gent getKB retract sense sense operators 7 public void move jTankWorldAdapter doMove public void attack jTankWorldAdapter doAttack 3 Get domain knowledge An agent can be viewed as being composed of two parts architecture and knowledge The agent architecture is fixed you should not change CAST architecture unless it is necessary Most of the agents behavior should be captured as knowledge We can group the knowledge part further into domain dependent knowledge and domain independent knowledge The domain independent knowledge is the knowledge that is needed to compliment certain missing features from architecture In the dTank team f
35. es with which agents can use to make decisions or to check the pre conditions of tasks in the PrT net The team structure and team plans in MALLET and coded domain knowledge will be parsed into a collection of PrT nets and an agent knowledge base Both the process and knowledge about team and domain are important mental states that are shared among the whole team Based on these shared mental states agents can anticipate the needs of other team members and demonstrate strong intelligent team assistance behaviors JARE JARE Java Automated Reasoning Engine is a back chaining theorem prover for making inferences using knowledge that is written in the form of a separate Horn clause knowledge base which stores agents beliefs Initially an agent has domain knowledge and belief about the team After the agent is launched in the domain environment the knowledge base is updated dynamically by newly acquired information through agents observations and communications Post conditions can also update the agents knowledge base with the effect of actions that were just carried out Furthermore an agent can update its own beliefs after making certain decisions JARE is mainly used to determine the truth value of conditions or constraints that need to be evaluated in interpreting MALLET expressions at run time JARE is also used to bind variables through queries for plan instantiation conditional execution and making communication decisions CA
36. ess states In the following sections we use a firefighting domain scenario to illustrate how the CAST monitor works for a firefighter team The CAST monitor includes the following tabs Monitor CAST Agents Agent F1 Agent F2 and Agent Am F1 F2 and Am are corresponding to firefighter agent 1 2 and an ambulance agent respectively 6 1 CAST team monitor Monitor CAST Agents is the panel for monitoring all individual agents behavior and status The functions of buttons are listed in table 1 Table 1 Buttons in the team monitor window Button Description Add Agent is reserved for dynamically adding agent to a team Next Step makes the agent proceed with one step Unpause unpauses all the agents to let them proceed automatically All the button will change to Pause All after it is pressed Cast Monitor Monitor CAST Agents Agent F1 AgentF2 Agent Am Current Agents Status of Agents Add Agent 19 52 10 F1 ask Am outOfControl fire 19 52 10 Am current message load 2 100 sender F1 type ask message outOfCont Next Step 19 52 10 F2 ask Am outOfControl fire Unpause All 19 52 10 Am current message load 1 100 sender F2 type ask message outOfCont S 19 51 51 Amiinitial plans fireFighting E agents 19 51 51 Am kernel starting 19 51 51 Am paused true 19 51 51 F2 initial plans fireFighting 19 51 51 F2 kernel starting 19 51 51 F2 paused
37. etermine who will take the action and synchronize their behavior if necessary During the act phase the kernel dynamically loads the class for the committed operation type creates an instance of it and invokes the operation In addition the kernel also applies the effects of the operation to its knowledge base and if needed proactively informs the new generated information to its teammates More specifically the kernel accomplishes the sense phase by invoking operations in the domain dependent library to obtain the needed information from the Domain Simulator The Domain Simulator determines which agent needs what information through analyzing relevant JARE rules and the specification of observabilities pre defined in MALLET programs As the Domain Simulator updates itself the actions executed by individual agents affect the world state it notifies the appropriate Domain Adapter which forwards the information to the corresponding kernels agents which then update their belief base maintained by JARE Engine to reflect the new acquired information See Figure 1 The decide portion is an area which exposes promising research topics in CAST agent architecture Amongst the features that have been developed and tested thus far are e Based upon the analysis of the MALLET programs every CAST agent uses the DIARG algorithm offline part discussed earlier for determining agent information needs relation and agent information production relation
38. he head For example consider the following rule which says that something is a good deal IF it has high quality and a low price good deal x quality x high price x low The antecedents are implicitly conjoined and ed together There is no way to encode disjunction except that multiple rules are alternatives to each other The head of each rule must be a positive literal but any of the other members of the clause antecedents may be negated Note that there are no quantifiers any variables are assumed to be universally quantified read forall It is of central importance that the head of each clause be a positive literal i e does not contain not In particular facts must always be positive literals Here is a brief list of commands you can run in Jare e query conjunction finds all solutions and prints out query with each set of bindings e query all conjunction same as query e query one conjunction finds first solution and prints out query with bindings substituted in e assert predicate adds fact to knowledge base e retract predicate retracts ALL clauses whose head unifies with given predicate may have constants and or variables gt query dog fido dog fido gt query cat fido fail gt query animal fido animal fido gt query animal a animal fido animal fifi animal tweety animal opus gt assert dog rex a
39. l accomplish their task To be expressive MALLET provides a rich set of constructs to define such procedures A process consists of invocations of operators or plans or arbitrary combinations using various constructs such as sequential parallel conditional or iterative blocks etc For example the following is a high level MALLET plan that characterizes the teamwork process for multiple teams to find and attack enemy base The plan is the second stage of an overall plan The first stage is to search and identify the location of the enemy base and the second stage is to attack the enemy base Therefore this plan includes the location of the enemy base as a precondition plan Attack Enemy Base pre cond is enemy homebase h1 at h1 x y 15 37 process par do scoutteam seeking do fighterteam attacking enemy do bomberteam attacking base x y endpar endproc endplan plan attacking enemy pre cond enemy e at e x y The keywords SEQ and PAR are used in MALLET to describe a sequential process and a parallel process respectively For instance the Attacked Enemy Base plan described above involves a parallel process with three branches These three branches describe subplans for three subteams The first branch specifies that the scout subteam should search for the enemy The second branch states that the fighterteam should attack enemies detected by the scouts The third branch specifies
40. ly different domains but also heterogeneous network environment 9 37 Agent communication Like most multi agent architectures messages from CAST agents are encoded in KQML Knowledge Query Manipulation Language 5 format that defines both the content of the messages and performatives of the communication CAST currently supports two kinds of performatives 1 information exchange performatives and 2 coordination performatives Proactive delivery of information needed by teammates is realized using a novel performative called ProInform The semantics of ProInform extend that of Inform to include the needs of addressee in the mutual belief the speaker attempts to establish with the addressee The semantics of performatives have direct impact on the conversation policy between agents in a team For instance a new type of reply can be introduced for ProInform such that the addressee can indicate that the information delivered is not needed or contrary to what the speaker believes We name the performative of this reply RejectNeed A more detailed discussion about the semantics of proactive communications among teammates can be found in 4 The second type of performatives is for agents to coordinate with each other regarding the execution of a team process For example at the conclusion of a team operation there will be communication by the performer s to let other members of the team in some cases not all members of a team nee
41. my tanks Once the team finds an enemy the team members communicate to inform each other of the enemy location and to cooperate for attack Next the team attacks the target together until the target is destroyed Such a process iterates until all the enemies have been destroyed Declarative domain knowledge includes moving directions stone locations and so on 5 37 3 About CAST 3 1 Overview CAST is a teamwork architecture that enables building collaborative agents as teams The agents can monitor the progress of team activities synchronize their behaviors and more importantly anticipate the needs of their teammates CAST uses a high level language called MALLET a Multi Agent Logic based Language for Encoding Teamwork to capture teamwork knowledge flexible enough for the team to adapt to a dynamic environment and efficient enough for the team to coordinate and assist with limited communications In addition each agent uses novel algorithms to 1 dynamically assign tasks to member of the team and 2 to anticipate information needs of teammates based on the shared knowledge 3 2 Main features The central aspect to our architecture is the maintenance of a Shared Mental Model SMM among the agents The SMM of a CAST agent has three components First team structure and team process knowledge is described in MALLET The structure knowledge describes roles in the team agents in the team and the role each agent can play Th
42. ng knowledge that is written in the form of a separate Horn clause knowledge base which stores agents beliefs Initially an agent has domain knowledge and belief about the team After the agent is launched in the domain environment the knowledge base is updated dynamically by newly acquired information through agents observations and communications Post conditions can also update the agents knowledge base with the effect of actions that were just carried out Furthermore an agent can update its own beliefs after making certain decisions JARE is mainly used to determine the truth value of conditions or constraints that need to be evaluated in interpreting MALLET expressions at run time JARE is also used to bind variables through queries for plan instantiation conditional execution and making communication decisions 5 1 Syntax of JARE language Jare uses a LISP oriented syntax with nested lists and infix notation where the first element in each list is often the name of a predicate and operator etc Capitalization and white spaces generally do not matter Comments are indicated by semi colons which cause disregard of all the remaining characters on the line The basic unit of expression in Jare is a predicate As is first order logic predicates have a predicate name followed by a list of arguments called terms However the predicate name is written inside the parentheses as the first member of the list Terms can eithe
43. nto the agent knowledge base becomes team members shared mental knowledge about the team which is important to help team agents to reason about their teammates for task allocation and proactive communications The main function of MALLET is to represent teamwork knowledge plans for a team The plans are organized in hierarchical structures such that each plan consists of one or several steps which are either sub plans or atomic operators Those steps are controlled by different types of processes which are specified as sequential parallel selective iterative or conditional The whole team plan is converted into a PrT net a process network which controls the execution of the team plan Pre conditions and post conditions are defined for each of the sub plans and operators Pre conditions specify the resource and information required for carrying out certain tasks or operations For 7 37 example it is a pre condition that an attack plan requires the target location before an agent can carry out the task The effect of the task is regarded as a post condition which in this example might be that the target is destroyed Later we will explain how the pre condition information is used for the agent to infer about information need of other teammates In addition to team design domain knowledge is also coded and loaded to the agent knowledge base Domain knowledge includes both facts about domain and inference rules as Horn claus
44. olling the process of CAST agent Util include some modules which are necessary for dealing with knowledge base 28 37 7 2 Basic steps of building a CAST team We adopt an agent oriented system analysis and design approach for building teams in CAST The main tasks of creating teams with CAST include knowledge engineering for agents and facilitating functions that handle connections between agents and a domain 1 Study domain and get rules APIs etc The first step is to study the domain A domain simulator defines the rules of the world such as the sensory visual audio inputs operation effects agents outputs and world state updates time resources As an agent designer you need to be clear about the environments which the agents will interact with Then you need learn about the interaction protocols that are defined as API functions or network messages For example the dTank defines its protocols as specified in table 3 By study the interactive protocols you should identify how to handle agents sensory inputs and apply agents operators In the dTank example we may identify that agents can sense information about a tank including id color etc We can also learn to apply operators such as move and turn Table 3 dTank communication protocol io input link dtank id lt lt string gt gt tank color lt lt string gt gt hit yes died yes status health lt lt int 1 10 gt gt
45. or example knowledge about how to aim a gun at an enemy is domain dependent knowledge on how to choose an indifference operator is domain independent We ae to capture agent s knowledge before the agents know how to collaborate in a omain The procedural knowledge contains a two phase plan search and attack as illustrated in Figure 8 Initially a team of agents wander around and search for enemy tanks Once an enemy is found the team members communicate to inform each other of the enemy location and to coordinate their attack Next team members attack the target together and destroy it This process iterates until all the enemies are destroyed Declarative domain knowledge includes moving directions stone locations and so on gt Move Y gt Turn Search Rotate Turret Lock target Y gt Aim Turret Attack gt Fire Figure 8 Agent plan decomposition 32 37 4 Teamwork knowledge In addition to domain knowledge the designer also needs to capture teamwork knowledge which defines the team composition roles and capabilities For example the following MALLET specifies a team for dTank domain agent F1 agent F2 team tankteam F1 F2 capability F1 move turn attack rotateTurret raiseShields scan capability F2 move turn attack rotateTurret raiseShields scan
46. oretic communication strategies The PrT net interpreter is used for interpreting manipulating the PrT nets so that all the team agents cooperate their behavior to embody the specified team behavior The internal PrT net representations of team processes are generated offline by MALLET Parser through compiling the team plans specified in MALLET Initially a team plan is only partially instantiated though fully represented in the PrT net As well as initiating appropriate communication actions at the appropriate time to evolve these partial plans determine parameters such as the actual doers of certain actions one of the tasks of NetIntepreter is to ensure all the team members to behave react and deliberate strictly according to the committed intended plans courses of actions and synchronize their behaviors whenever necessary At each computation step CAST agents sequentially execute three mental actions sense decide act During the sense phase each individual agent queries the world simulation server to update its knowledge of the current state of the world check its message queue and process the messages from other agents in this round During the decide phase by using NetIntepreter each agent examines the PrT net representation of the team plan to see if there are any pending actions for which it is responsible In cases of ambiguity about the actual doers of some pending actions the agents might have to communicate in order to d
47. r be constants symbols or numbers variables or functions Variables are indicated by symbols prefixed with a such as x or temp Functions are like predicates in that they are lists with function names and arguments though they occur as arguments inside other functions or predicates Here are some examples e teaches bill ai course e has phd bill e sister judy sis e loves jude mother of judy loves is a predicate mother of is a function Predicates can also be negated by enclosing them in another list whose first element is not Predicates and negations of predicates together are called literals Predicates without negations are sometimes called positive literals while negated predicates are sometimes referred to as negative literals not warm december a negative literal Sentences in Jare are based on Horn clauses which are made out of one or more predicates When a Horn clause contains a single literal it acts as a fact note it has an extra set of enclosing parentheses Facts by themselves cannot be negated Here are examples of two facts 18 37 warm july cold december When 2 or more literals are included the Horn clauses can be read as a rule Specifically the first literal is known as the head or consequent of the rule and remaining literals are called the body or antecedents A rule can be understood by reading in a backwards way by saying if after t
48. r not this agent each agent does such an analysis is to perform the chosen friable next action In short during the decide phase each agent determines what operation if any it should perform during its next cycle performing if necessary dynamic agents assignment and determines what communication should take place with which other teammates 13 37 4 MALLET MALLET Multi Agent Logic Language for Encoding Teamwork is a logic based language for specifying the structures and processes of agent teams MALLET syntax is based loosely on LISP in the sense of using s expressions and prefix notation Variables are indicated with prefix The full syntax for MALLET is shown in Appendix A 4 1 Team structure knowledge At the top level MALLET allows expression of knowledge about team structure in terms of e Agent role relationship i e which agent plays which role e Membership of a team For example the following specification defines the team attackteam which has two members FO and UO playing different roles i e fighter and bomber respectively plays role FO fighter plays role UO bomber team attackteam FO UO The actions agents are able to perform are defined in terms of operators Operators specified by preconditions and effects are classified into individual and team operators Individual operators are executed by only one agent at a time whereas team operators are performed by a set of agent
49. recondition successfully 2 After the number of synchronization messages received by an actor exceeds or equal to the minimum number of agents required to performing the operator as specified by the num constraint of the team operator in MALLET the actor performs the team action Similar synchronization scheme is also used for agents to synchronize in invoking a team plan using JOINTDO This simple synchronization scheme does not consider communication delays Alternative synchronization scheme will be needed if agents need to synchronize in an environment that communication delays are not negligible Distributed coordinations of a team plan are realized by coordination messages between members of the team Each coordination message contains an identifier that refers to a specific task i e operator or subplan in a team plan We describe the communication protocol for achieving coordination below e Ifan agent is not an actor for the current task i e operator of plan in the team plan the agent waits for coordination messages associated with the task from teammates e Ifthe agent is an actor for the current task it starts performing the task synchronize with teammates if needed e When the agent finishes the task it sends a coordination message containing the identifier of the task to all other teammates involved in the team plan e When an agent waiting for a coordination message receives the message it proceeds to
50. s e g those playing a given role For example the following specification defines individual operator move ioper move dir pre cond at self x y can_move dir new x new y effects at self new x new y not at self x y This operator describes an action for the agent to move a step along a direction The precondition of the operator includes two predicates connected implicitly by logic conjunction operator 1 e AND The first predicate retrieves the current location of the agent referred in MALLET by the keyword self The second predicate tests whether the agent can move in the direction given and if so binds the values of such a move to the variables new x and new y The evaluation of the test involves Horn clauses in JARE that considers obstacles and the boundary of the world in which the agent navigates After the move the operator s effect updates the location of the agent As a second example the following defines a team operator co_fire which involves at least two agents firing at a given coordinate specified by the arguments x and y simultaneously toper co_fire x y num ge 2 14 37 The num constraint specifies the number of agents that should be involved in the team operator Obviously agents involved in a team operator need to be synchronized This is achieved by the CAST kernel The capabilities of individual agents can be specified either explicitly via a capability
51. sserting dog rex gt assert dog poochie asserting dog poochie 19 37 gt retract bird x retracting bird x gt query animal a animal fido animal fifi animal rex animal poochie 5 2 List of features In this section several important features implemented in Jare are discussed The first is negation While technically Horn clauses are not supposed to contain negative literals Jare allows antecedents in a rule to be negated Syntactically a negative literal is written by enclosing a predicate in another list with not as the prefix an example is shown above The semantics of negation is handled exactly like Prolog through negation as failure That is an antecedent not P is found to be true exactly when P cannot be proved Jare recursively starts a new search for a proof of P and then continues the original proof if and only if no solution for P is found Note that the proof of P is carried out in the current binding environment with current substitutions but the fact that the original proof proceeds only if the proof of P fails means that the binding environment itself will not be modified As an example here is a rule with a negative antecedent It says that something is a week day if it is a day but not a weekend Days could be 7 facts enumerating Sat Sun Mon Each will be tried but those that satisfy weekend i e Sat and Sun will be filtered out as
52. that a team of bomber agents should attack the enemy base The precondition of the attacking enemy plan includes the location of enemy i e at e x y Therefore a scout agent will proactively deliver the location of detected enemy to the fighterteam Other types of MALLET processes include conditional i e IF statement iteration 1 e WHILE statement selection 1 e CHOICE statement Components of these processes are other MALLET process Hence the MALLET language allows any processes to be embedded in another process to form more complex processes An important feature of MALLET is its flexibility in specifying the actors of a process using DO and JOINTDO We describe each of these MALLET constructs below The formal BNF style notation of the MALLET process is as follows lt MalletProcess gt lt OperatorInvocation gt lt PlanInvocation gt SEQ lt MalletProcess gt PAR lt MalletProcess gt IF lt COND gt lt Cond gt lt MalletProcess gt lt MalletProcess gt WHILE lt COND gt lt Cond gt lt MalletProcess gt FOREACH lt COND gt lt Cond gt lt MalletProcess gt FORALL lt COND gt lt Cond gt lt MalletProcess gt 16 37 CHOICE lt MalletProcess gt JOINTDO lt AND gt lt OR gt lt XOR gt lt ByWhom gt lt MalletProcess gt DO lt ByWhom gt lt MalletProcess gt AGENTBIND lt AgentVars gt
53. true 19 51 51 F1 initial plans fireFighting 19 51 51 F1 kernel starting 19 51 51 F1 paused true Figure 2 CAST team monitor 23 37 6 2 Individual agent monitor Under the individual agent tab the information and control panels of the specific agent is displayed The behaviors and status of this agent are displayed in the window on the left side The buttons and their functions are listed in Table 2 Table 2 Buttons in the individual monitor window Buttons Description Unpause Pause unpauses this agent only after unpaused the button changes to Agent Pause Agent Show Hide Process generates a tab for every plan of this agent and under every tab Nets is the process net for the plan after it is pressed the button changes to Hide Process Nets Show MALLET shows the MALLET specification in the window of Knowledge Specifications Base Output Save as MALLET saves the MALLET specifications in the window of Knowledge File Base Output as a MALLET File Cast Monitor Monitor CAST Agents Agent F1 AgentF2 Agent Am Agent Characteristics Specs Agent Name F1 Simulation Domain cast3 domain rescueworld RescueDomain Show MALLET Specifications Save as MALLET File cast3 util Jare JareKBW rapper wey Knowledge Base Output Interactions Receiver ID f4 Message Type control Message Text
Download Pdf Manuals
Related Search