CArtAgO: A Framework for Prototyping Artifact-Based Environments in MAS

Alessandro Ricci, Mirko Viroli, and Andrea Omicini

Alma Mater StudiorumUniversit` a di Bologna via Venezia 52, 47023 Cesena, Italy {a.ricci,mirko.viroli,andrea.omicini}@unibo.it

Abstract. This paper describes CArtAgO, a framework for developing artifact-based working environments for multiagent systems (MAS). The framework is based on the notion of artifact, as a basic abstraction to model and engineer objects, resources and tools designed to be used and manipulated by agents at run-time to support their working activities, in particular the cooperative ones. CArtAgO enables MAS engineers to design and develop suitable artifacts, and to extend existing agent platforms with the possibility to create artifact-based working environments, programming agents to exploit them. In this paper, rst the abstract model and architecture of CArtAgO is described, then a rst Java-based prototype technology is discussed.

Introduction

Artifacts have been recently proposed as rst-class abstractions to model and engineer agent working environments in software MAS (multiagent systems) [1]. The background view, shared with other recent approaches in MAS literature see [2,3] for a survey, is that the environment plays a fundamental role in engineering of MAS. On the one hand, environment is a suitable locus for engineers to embed responsibilities, impacting on MAS design and development; on the other hand, it is a source of structures and services that agents can suitably use at run-time to support and improve their activitiesboth individual and social ones. The specic notion of working environment is intentionally analogous to the notion of human cooperative working environments, as they are studied by disciplines and theories in human science, such as Activity Theory and Distributed Cognition, and recently adopted also in the context of CSCW (Computer Supported Cooperative Work) and HCI (Human-Computer Interaction) [4,5]. There, a working environmentalso referred as eld of work is such part of the environment explicitly designed to support and realise agent working activities. Typically, it is modelled as set of objects, tools, more generally artifacts, which are constructed, shared, and either cooperatively (or competitively) used by humans, so as to mediate and sustain their activities. Analogously to human society, such a perspective is likely to be fundamental also in the context of agent societies, in particular for designing and programming
D. Weyns, H.V.D. Parunak, and F. Michel (Eds.): E4MAS 2006, LNAI 4389, pp. 6786, 2007. c Springer-Verlag Berlin Heidelberg 2007

68

A. Ricci, M. Viroli, and A. Omicini

complex software systems based on MAS. Given that MAS are growing increasingly complex, one may easily foresee that the next step in the evolution of cognitive MAS will require MAS models and architectures to deal with agents situated within suitable working environments. There, agents would autonomously construct, share, and co-operatively use dierent kinds of artifactdesigned either by MAS designers or by the agents themselvesto perform MAS activities. It is worth noticing that such a perspective shares the aims and principles developed by the research work in Distributed Articial Intelligence about theories of interaction, environments and the role of tools [6,7]. The artifact abstraction is at the heart of this conceptual frameworkwhich can be referred as A&A (agents and artifacts)and promotes a methodology for modelling and engineering working environments, by introducing new concepts and elements that impact on system design, development and run-time management. Artifacts can be generally conceived as passive, function-oriented computational entities, explicitly designed to provide some kind of function, and then to be used by agents to support their individual and collective (social) activities [1]. The notion of function here refers to the meaning that is generally used in human sciences such as sociology and anthropology, as well as in some recent work in AI [7], that is, the purpose for which the object has been designed forfor an artifact, to support agent activities. This view directly impacts on the foundation of interaction and activity in agency: a MAS is conceived as an (open) set of agents that develop their activities by (i) computing, (ii) communicating with each other, and (iii) using and possibly constructing shared artifacts. Artifacts could be either the targets (outcome) of agent activities, or the tools that agents use as means to support such activities: as such, they are useful to reduce complexity of task execution. For instance, coordination artifacts [8] are artifacts providing coordination functionalitiessuch as blackboards, tuple spaces or workow engines. In this paper we introduce and discuss CArtAgO (Common Artifacts for Agents Open framework), a framework for prototyping MAS applications with artifact-based working environments. Essentially, CArtAgO provides (i) the API to dene any useful kind of artifacts, (ii) the API to be exploited by agents (agent programmers) for interacting with working environments populated by artifactsin particular to instantiate, use, manipulate artifacts, and (iii) a run-time environment supporting the existence and dynamic management of working environments. CArtAgO does not introduce any specic model or architecture for agents and agent societies: the framework is meant to be integrated and used with existing agent platforms, possibly characterised by heterogeneous kinds of agent architectures. From a conceptual point of view, CArtAgO makes it possible to build MAS composed by heterogeneous agent societies, made of reactive and cognitive agents programmed with dierent agent languages or architectures, sharing the same working environments, and interacting through suitable mediating artifactsbesides communicating via ACL as usual.

CArtAgO: A Framework for Prototyping Artifact-Based Environments

69

The rest of the paper is organised as follows: rst, we describe the abstract model and architecture of CArtAgO (Sect. 2), focusing in particular on the core of API introduced by the framework (Sect. 3); then, we describe a rst concrete implementation prototype (Sect. 4) developed in Java, implementing the core part of the abstract model previously dened.

CArtAgO Abstract Model and Architecture

In this section we describe the basic elements and structure of CArtAgO working environments, by taking as a reference the abstract architecture schema described in [3] and depicted in Fig. 1, useful to understand CArtAgO with respect to the other approaches. Accordingly, the abstract architecture of CArtAgO (and of CArtAgO working environments) is composed by three main building blocks (see Fig. 2): (i) agent bodies as the entities that make is possible to situate agents inside the working environment; (ii) artifacts as the basic building blocks to structure the working environment; and (iii) workspaces as the logical containers of artifacts, useful to dene the topology of the working environment.
MAS Application
Application Agents

Application Environment Application Specific Logic

Execution Platform

Agent Middleware / Infrastructures

Action Perception

Environment Middlewares / Infrastructures

MAS Middleware Layer

SW

Operating Systems, Virtual Machines & Other Middlewares Deployment

Context

Physical Support

Hardware & Network Physical World

HW Deployment Context

Fig. 1. Abstract representation of MAS layers with environment-based supports as depicted in [3]. Rectangles represent S/W and H/W tiers of the application at dierent levels. Agents are expressed as circles, environment abstractions as boxes. Arrows from agents to environment abstractions represent actions, dashed arrows in the opposite direction represent perceptions. Arrows between agents represent direct agent communications, while arrows between environment abstractions represent intra-environment interactions. Vertical lines represent the infrastructure supporting a concept at the MAS application level.

70

A. Ricci, M. Viroli, and A. Omicini

Application Agents Artifact-based working environments
workflow engine blackboard

MAS Application

shared kb

map

Application Specific Logic

Agent Framworks / Middlewares

artifacts

MAS Middleware Layer

Execution Platform

agent bodies

workspaces

CARTAGO

Any OS

JVM

JVM OS

Fig. 2. MAS layers adopting CArtAgO support. Application environments are modelled in terms of artifact-based working environments. The CArtAgO middleware manage the life-cycle of working environments, composed by artifacts grouped in workspaces. Agent bodies are used to situate agents inside the working environments, executing actions upon artifact and perceiving artifacts observable state and events.

2.1

Agent Bodies

Agent bodies are what actually enable the coupling between an agent (mind) and a CArtAgO working environment. For each agent aiming at working inside a CArtAgO environment, an agent body is created. The agent body contains eectors to perform actions upon the working environment, and a dynamic set of sensors to collect stimuli from the working environment. The agent body is meant to be controlled by the agent, which actually plays the role of the pilot of the body. For the purpose, the agent body exposes a controlling interface that the agent mind could suitably exploit to interact with the environment. By piloting their agent bodies, agents can interact with their working environment, executing actions provided for artifact construction, selection and usage, and perceiving observable events generated from such artifacts. Dierently from the approach typically adopted in traditional agent architectures and more similarly to active perception [9], here perception is modelled as an intentional action referred as sensing. More precisely, environment observable eventsgenerated by artifactsare collected as stimuli by sensors which are part of the agent body. An agent can dynamically and exibly link and unlink to its body dierent kinds of sensor, with dierent functionalities, such as buering, ltering, ordering, and managing priorities. So, in CArtAgO sensing is the internal action that agents execute on their sensors to become aware (perceive) of the stimuli collected by the sensors. Stimuli typically concern observable events generated by artifacts.

CArtAgO: A Framework for Prototyping Artifact-Based Environments

71

2.2

Artifacts

Artifacts are the basic bricks managed by CArtAgO framework. Each artifact has a logic name specied by the artifact creator at instantiation-time, and an id, released by the framework, to univocally identify the artifact. The logic name is an agile way for agents to refer and speak about (shared) artifacts, while the id is required to identify artifacts when executing actions on them. The full name of an artifact includes also the name of the workspace(s) where it is logically located. Since an artifact can be located in multiple workspaces, the same artifact can be referenced by multiple full names. Usage Interface & Observable Events. Analogously to artifacts in our society, the basic model which characterises the interaction between agents and artifacts is based on a notion of use and observation. Agents can use an artifact by triggering the execution of operations listed in the artifact usage interface. An operation is characterised by a name and a set of typed parameters. The execution of an operation typically causes the update of the internal state of an artifact, and possibly the generation of one or multiple observable events including error conditions that can be possibly collected by agents sensor as they are generated, and perceived by agents through explicit sensing actions. The usage interface of an artifact can change according to artifact observable state, exposing dierent sets of operations according to the specic functioning state of the artifact. The notion of observable state is adopted to structure the functioning behaviour of an artifact in a set of labelled states, which can be recognised (observed) by the artifact users. For each artifact type a nite set of labelled observable states can be dened. For each concrete instance, the notion of current observable state is dened, and its value can change dynamically, during artifact functioning. Then, for each observable state a dierent usage interface can be dened. This feature makes it possible to structure the overall usage interface of an artifact, providing the right interface according to the functioning stage of the artifact. In other words, an artifact can expose dierent set of operations according to its observable state. Dynamically, an agent can trigger the execution of an operation on an artifact if and only if the operation is (currently) part of the usage interface; if the operation does not belong to the usage interface, the agent action fails. Function Description and Operating Instructions. In order to support a rational exploitation of artifacts by intelligent agents, each artifact is equipped with a function description, i.e. an explicit description of the functionalities it provides, and operating instructions, i.e. an explicit description of how to use the artifact to get its functionfor instance in terms of the usage protocols that the artifact support. These descriptions are meant to be useful for cognitive agents thatby suitably inspecting and interpreting themcan (i) dynamically reason about which artifacts can be selected to support their activities, and (ii) get instructions to support activity execution, making it easier to set up plans and to reason about the expectation of using an artifacts. We consider such issues of foremost importance, at the core of the notion of computational environments

72

A. Ricci, M. Viroli, and A. Omicini

designed to support the activities of agentsin particular cognitive / rational agents. Actually, research on these aspectsin particular on formal models and languages that can be used to specify function description and operating instructions, and their injection in existing agent reasoning architectures (such as BDI)is still to be fully developed: we forward the interested reader to [10] for the rst results. In CArtAgO, we provide a minimal enabling support to such issues, by modelling function description and operating instructions as at strings, specied by artifact designers and dynamically inspectable (observable) by agents through suitable actions. Currently, there is no predened syntax and semantics for such information (see future work for comments on this point). 2.3 Workspaces

Artifacts are logically located within workspaces, which can be used to dene the topology of the working environment. A workspace can be dened as an open set of artifacts and agents creating and using them: artifacts can be dynamically added to or removed from workspaces, agents can dynamically enter (join) or exit workspaces. The same artifact can belong to multiple workspaces. In CArtAgO, each workspace is created by specifying a logic name and is univocally identied by an id. By dening a topology of the environment, workspaces make it possible to structure agents and artifacts organisation and interaction, in particular functioning as scopes for event generation and perception, and artifact access and use. On the one side, a necessary condition for an agent to use an artifact is that it must exist in a workspace where the agent is located. On the other side, events generated by the artifacts of a workspace can be observed only by agents belonging to the same workspace. Intersection and nesting of workspaces are supported to make it possible to create articulated topologies. In particular, intersection is supported by allowing the same artifacts and agents to belong to dierent workspaces.

Core Primitives

After providing an overview of CArtAgO main components, in this section we describe the basic abstract set of core API provided by the framework on the one side to be used by agents (or agent programmers dening agent behaviour) to interact within working environments, and on the other side for dening artifact types, that is programming artifacts behaviour. 3.1 Agent Side

On the agent side, the API is represented by a set of primitives to control agent bodies and that eventually result in executing actions inside the working environment, making it possible basically to create and use artifacts, and perceive artifact observable state and events. Table 1 provides an abstract description of such primitives, grouped according to their functionalities:

CArtAgO: A Framework for Prototyping Artifact-Based Environments Table 1. Actions available to agents to manage artifacts and workspaces
createArtifact(Name,Template,Config,{WsID}):ArID disposeArtifact(ArID) getArtifactID(Name,{WsID}):ArID execOp(ArID,OpName,{Args},{SensorID}) sense({SensorID},{Pattern},{Timeout}):Perception focus(ArID,SensorID) unfocus(ArID,SensorID)

73

Artifact construction and disposal

Artifact selection & use

Artifacts inspection

getFD(ArID): FDDescr getOI(ArID): OIDescr getUID(ArID): UIDDescr getState(ArID): StateDescr

Sensor management

linkSensor(SensorType,SensorConfig): SensorID unlinkSensor({SensorID})

Workspaces management

getWsID(WsName):WsID createWS(WsName):WsID disposeWS(WsID) registerArtifact(ArID,WsID) deregisterArtifact(ArID,WsID) joinWS(WsID) exitWS(WsID)

Artifacts construction & disposal Basic primitives are provided to create (createArtifact) and dispose (disposeArtifact) artifacts dynamically. To create an artifact, a logic name must be specied, along with the Template that identies the type of the artifact to be created, the initial conguration parameters needed for artifact creation and optionally the workspace where the artifact should be created. The action can fail if the template is unknown or the artifact instantiation is not completed due to some kind of problem (e.g. wrong initial conguration). Artifact discovery & use These primitives constitute the core of agent / artifact interactions, enabling an agent to use an artifact by executing operations and observing artifact state and events. To execute an operation, the action execOp is provided, specifying the artifact identier, the operation name, the parameters, and (optionally) the specic sensor where to collect observable events generated by the artifact as a consequence of the operation execution. The action can fail either because the specied artifact is not available, or because the operation cannot be executed since it is not part of artifact usage interface. Action success means that the execution of the specied operation has been successfully triggered. The identier of an existing artifact can be obtained by the getArtifactID primitive, specifying the artifact name and (possibly) its location (workspace). After triggering the operation, an agent can observe related events through codesense primitives on the sensor specied in execOp. By executing sense

74

A. Ricci, M. Viroli, and A. Omicini

actions, an agent is made aware of the stimuli that are dynamically collected by a sensor. In particular, the eect of the action is to fetch (remove) a stimulus from the sensor and to return it to the agent as a perception. The action fails if no stimuli are available. Dierent types of sensors can provide dierent semantics establishing the order in which events are fetched. A time parameter can be optionally provided to indicate the duration for the sensing action: if no events are available in the sensor within the specied time-frame, the action fails. By default, the time-frame is zero. In order to support forms of data-driven (or, equivalently, lter-driven ) sensing, a pattern parameter can be specied acting as a lter for fetching (selecting) the perception. Conceptually, the pattern denes a set of perceptions: a perception is fetched if and only if is included in the set. Typically, the pattern can be represented by a Boolean function, establishinggiven a perceptionif either it is part or not of such a set. It is worth noting that specifying a pattern in a sense action is dierent from creating sensors that lter stimuli as they are collected. An available stimulus which is not fetched by a sense action because not satisfying the pattern is not removed from the sensor and can be possibly fetched by subsequent sense actions. Finally, primitives for continuous observation are provided: by executing a focus action, an agent becomes a permanent observer of the artifact whose identier is specied as a parameter. As a permanent observer, all the observable events generated by the artifact are automatically collected by the sensor specied as second parameter, as they are generated; unfocus stops the observation. Artifacts inspection In order to support a cognitive use of artifacts, a basic set of primitive is provided to inspect the function description (getFD), the operating instructions (getOI), the usage interface (getUID), and the dynamic observable (exposed) state (getState) of the artifact. Sensor management Two basic primitives are provided to dynamically link and unlink sensors to the agent body: linkSensor links a sensor of the specied type and conguration to the body, returning the identier to be used to refer the sensor; unlinkSensor unlinks a previously linked sensor. Workspace manipulation Finally, a basic set of primitives is provided to manipulate the logical topology of the environment, modelled through workspaces. Such primitives range from joinWS and exitWS to join and leave a workspace, to getWsID for getting a workspace identier given its name, createWS for directly creating a new workspace and disposeWS for completely removing a workspace. Since the same artifact can be part of multiple workspaces, some basic primitives are provided to register (registerArtifact) / de-register (deregisterArtifact) an artifact in / from a workspace, specifying the workspace id. Most of these core services have been implemented in the prototype described in Sect. 4.

CArtAgO: A Framework for Prototyping Artifact-Based Environments Table 2. Basic primitives for artifact programming
Observable event generation genEvent({OpID},EventType,{EventContent}) genEventInWsp(EventType,{EventContent})

75

Operation management

getOpID: OpID

Observable state management

setObservableState(ObsStateName) getObservableState:ObsStateName

3.2

Artifact Side

On the artifact side, CArtAgO provides a support to dene new types of artifact, dening artifact structure and behaviour. The specic programming model adopted to implement in Java artifact types is described in detail in Sect. 4. Here we report the basic set of abstract primitives which can be exploited when dening artifact behaviour (see Table 2), useful essentially for generating observable events and switching artifact observable state. Observable events can be generated as either related or not to the specic execution instance of an operation. For the purpose, each operation triggered on the artifact is labelled by a unique operation identier (type OpId in the tables). Such an identier can be explicitly retrieved by the getOpId primitive during the execution of the operation (as part of its execution body). Operation identiers are meant to be manageable as normal data structures, for instance, creating list of operation identiers and then generating events related to these operation when necessary, during artifact functioning, across operation executions (this aspect will be claried by a concrete example described in Subsection 4.2). An event can be then generated using the genEvent primitive by specifying the operation identier to which the event must be related, as observable eect of this operation (and of the agent action that caused it). If no OpId is specied, the event is considered related to the current operation triggered. The eect of the execution of these primitives is the generation of an event which is eventually collected by the sensor (if specied) of the agent that triggered the operation and by all the agents that are observingvia focusthe artifact. To generate event unrelated to a specic operation execution, the primitive getEventInWsp is provided, which generates an event which is observed by all the agents focusing on the artifact. Finally, a couple of primitives are provided to manage the current observable state of the artifact, in particular to set a new value with setObservableState, specifying a label identifying one of the possible set of observable states dened by the artifact type, and to retrieve current value with getObservableState. Besides the events explicitly generated with the genEvent primitive, some other kinds of event are automatically generated by the framework and made observable to agents interacting with an artifact. In particular, an event is generated whenever the execution of an operation is completed, and whenever a new

76

A. Ricci, M. Viroli, and A. Omicini

observable state is set (details about the specic types of these events are provided when describing in next section).

A First Prototype

A rst prototype implementing most of the functionalities described in the previous section has been developed in Java and is available for download at the CArtAgO project web site1 . Our objective was to set up a rst framework for prototyping and experimenting applications engineered upon the A&A metamodel, and so designed in terms of set of agentspossibly with heterogeneous models and architecturessituated in the same working environment, designed in terms of specic kind of artifacts. The framework itself is not meant to dene or constrain the specic agent architecture adopted to dene the behaviour of the individual agents: conversely, the framework is meant to be integrated and exploited with external agent frameworks or platforms, in particular with those that adopt Java as underlying implementation language, extending them so as to support the creation and use of artifact-based environment according to the A&A perspective. As an example, simpA (simple A&A programming environment) is a fulledged agent-oriented framework for prototyping general-purpose applications based on CArtAgO. Basically, simpA provides a support for developing MAS based on agents with an activity-oriented architecture, with a native support for creating and using artifact-based working environment, engineered upon CArtAgO. The interested reader is invite to refer to the simpA web site2 . Based upon CArtAgO and simpA, simpA-WS is a framework for prototyping service-oriented applicationin particular Web Service-basedin terms of agents and artifacts. There, artifacts are used on the client side as interfaces for user application agents to exibly access and use Web Services, on the service side as interfaces for service agents to get Web-service messages and requests to be processed, and to provide responses. More information can be found at the simpA-WS web site3 . Working in the rst real-world application examples, simpA-WS is currently being investigated as an agent-based technology for prototyping service-oriented applications in the context of logistics 4 . 4.1 Prototype Overview

The framework is composed by four main parts: API for setup working environments The entry point class of the framework is the Cartago class (sketched in Fig. 3), which mainly provides static services to create or get the reference to existing working environments
1 2 3 4

CArtAgO web site: http://www.alice.unibo.it/projects/cartago simpA web site: http://www.alice.unibo.it/projects/simpa simpA-WS web site: http://www.alice.unibo.it/projects/simpaws http://www.alice.unibo.it/projects/a4stil

CArtAgO: A Framework for Prototyping Artifact-Based Environments

public class Cartago { public static synchronized ICartagoEnvironment getInstance(String name){...} public static synchronized ICartagoEnvironment getInstance(String name, ICartagoLoggerManager logger){...} public static String getVersion(){...} }

77

Fig. 3. Entry point class for the CArtAgO framework. The class can be used to instantiate and get the reference to working environment.
public interface ICartagoEnvironment { IAgentBody getAgentBody(String name) throws AgentBodyAlreadyPresentException; ArtifactId createArtifact(String name, Class template, ArtifactConfig param) throws ArtifactAlreadyPresentException, UnknownArtifactTemplateException, ArtifactConfigurationFailedException; ArtifactId createArtifact(String name, Class template) throws ArtifactAlreadyPresentException, UnknownArtifactTemplateException, ArtifactConfigurationFailedException; void registerLogger(ICartagoLogger logger); void unregisterLogger(ICartagoLogger logger); }

Fig. 4. Interface for working environments, providing services for creating agent bodies, and for directly creating artifacts, useful to setup the initial conguration of the environment

identied by a logic name. Once created or retrieved the reference to a working environment, it is possible to use the services provided by its interface ICartagoEnvironment, sketched in Fig. 4to setup the environment possibly creating an initial set of artifacts (besides the ones created dynamically by agents), and to create agent bodies, for enabling agents participation to the environment. API for controlling agent bodies From the agent point of view, the participation and interaction within a working environment takes place through an agent body. The creation of an agent body is provided as the getAgentBody provided by a working environment. Such a creation is typically done during agent initialisation. Once its agent body is created inside the environment, the agenthere conceived as the agent mindcan control it by suitably exploiting the IAgentBody interface implemented by the agent body, containing the core set of API described in Subsection 3.1. A sketch of the IAgentBody interface is reported in Fig. 5. It is possible to recognise the primitives for creating and disposing artifacts, for executing operations, sensing perceptions, managing sensors, and so on. API for dening artifact types A core part of the framework is given by the support provided to dene new kind of artifacts, programming their structure and behaviour. We adopted a programming model that favours rapid prototyping of artifacts, exploiting as much as possible the support

78

A. Ricci, M. Viroli, and A. Omicini

public interface IAgentBody { ArtifactId createArtifact(String name, Class template, ArtifactConfig param) throws ArtifactAlreadyPresentException, UnknownArtifactTemplateException, ArtifactConfigurationFailedException; ArtifactId createArtifact(String name, Class template) throws ArtifactAlreadyPresentException, UnknownArtifactTemplateException, ArtifactConfigurationFailedException; void disposeArtifact(ArtifactId id) throws UnknownArtifactException; ArtifactId getArtifactId(String name) throws UnknownArtifactException; OpId execOp(ArtifactId id, Op op) throws OperationException; OpId execOp(ArtifactId id, Op op, SensorId sid) throws OperationException; Perception sense(SensorId sensorId) throws NoPerceptionException; Perception sense(SensorId sensorId, IPerceptionFilter p) throws NoPerceptionException; Perception sense(SensorId sensorId, int dt) throws InterruptedException, NoPerceptionException; Perception sense(SensorId sensorId, IPerceptionFilter p, int dt) throws InterruptedException, NoPerceptionException; void focus(ArtifactId aid, SensorId sid) throws SensorNotLinkedException; void unfocus(ArtifactId aid); SensorId linkSensor(AbstractSensor s); void unlinkSensor(SensorId id) throws CartagoException; }

Fig. 5. Interface to control an agent body, including methods for triggering the execution of agent actions for artifact creation (createArtifact), artifact disposal (disposeArtifact), artifact discovery (getArtifactId), for triggering the execution of operation (execOp), for sensing perceptions (sense), for continuously observing artifacts (focus, unfocus), and for managing sensors (linkSensor,unlinkSensor)
public abstract class Artifact { ... protected final ArtifactId getId(){...} protected final OpId getOpId(){...} protected final OpRequestDescriptor getOpRequestDescriptor(){...} protected final void genEvent(String type) {...} protected final void genEvent(String type, Object content) {...} protected final void genEvent(OpId id, String type) throws InvalidOpIdException {...} protected final void genEvent(OpId id, String type, Object content) throws InvalidOpIdException {...} protected final void genEventInWsp(String type, Object content) {...} protected final void setObservableState(String state) throws UnknownArtifactStateException {...} protected final String getObservableState(){...} }

Fig. 6. Base abstract class to dene new artifact types. The basic set of primitives useful for programming artifact observable behaviour (in particular to generate observable events, to set and retrieve the observable state) are implemented as protected methods of this class.

CArtAgO: A Framework for Prototyping Artifact-Based Environments

79

given by the Java object-oriented environment. Accordingly, an artifact type can be dened by extending the basic Artifact class provided in the API: at run-time, artifacts instances are instances of this class. A sketch of the base class is shown in Fig. 6: the core set of the primitives described in Subsection 3.2 are available as protected methods provided by the class. The artifact internal state is dened in terms of instance elds of the class, and the behaviour of operations can be dened by suitable instance methods of the class. In particular an operation Op(Params) can be implemented by a method of the kind:
@OPERATION(State1,State2,...) void Op(Params){...}

The annotation @OPERATION5 is used to explicitly state that what follows is not to be interpreted as a normal method (meant to be invoked by other objects) but rather as the body of an artifact operation. It is worth remarking that methods representing operations have no return argumenta return argument would be meaningless in CArtAgO abstract model, as well as in the A&A general meta-model. Currently, the concurrency model adopted for artifacts prevents operation execution requests to be served sequentially, so that only one operation at a time can be in execution on an artifact. Such a choice is quite eective in avoiding basic problems related to concurrent use of artifacts by agents (and in particular concurrent updates of artifact internal state). At the same time, this choice limits quite strongly the concurrency in artifact use, so future work will be devoted to explore further this issue. As depicted in Fig. 6 and described in Subsection 3.2, observable events can be generated in the body of an operation by a family of primitives of the kind genEvent, specifying the event type, optionally an event content and the operation identier to which the event must be related (OpId parameter). Events are collected by agent body sensors as stimuli, and then perceived by agents through sense action. Fig. 6 also includes the primitives that can be used to set and retrieve the current observable state of the artifact (setObservableState and getObservableState, respectively). The manual of the artifact, containing information about function description, the operating instructions, as well as the list of the observable states, can be explicitly declared through the @ARTIFACT MANUAL annotation preceding the artifact class declaration. If no states are declared, a single default state is dened. Dened the list of the observable states, an artifact programmer can specify the shape of the usage interface in relationship to the artifact observable state. This is possible by explicitly stating in the annotation of an operation what are the observable states in which the operation is meant to be visible (specifying @OPERATION({State1,State2,...})). If an operation has no states declared, then the operation is meant to be visible in all the states. As a simple example, Fig. 7 shows the denition of an artifact type called MyArtifact (on the left), and an example of artifact use by an agent (on
5

Annotations have been introduced along with the 5.0 version of Java.

80

A. Ricci, M. Viroli, and A. Omicini

\begin{verbatim} @ARTIFACT_MANUAL( states = {"stateA","stateB"}, start_state = "stateA", oi = @OPERATING_INSTRUCTIONS("..."), fd = @FUNCTION_DESCRIPTION("...") ) public class MyArtifact extends Artifact { private int count; private int max; public CounterArtifact(int max){ this.max = max; count = 0; } @OPERATION({"stateA"}) void op1() { count++; genEvent("new_value",count); if (count >= max){ setObservableState("stateB"); } } @OPERATION({"stateA","stateB"}) void op2() { genEvent("value",count); } }

... ICartagoEnvironment env = Cartago.getInstance("..."); IAgentBody myBody = env.getAgentBody("..."); ArtifactId aid = myBody.getArtifactId("myArtifact"); SensorId sid = myBody.linkSensor(new DefaultSensor()); BasicFilter myFilter1 = new BasicFilter({"new_value"}); BasicFilter myFilter2 = new BasicFilter({"op_completed", "state_changed"}); boolean state_changed = false; while (!state_changed){ try { myBody.execOp(aid,"op1",sid); // operation triggered: // sensing for one second for new_value events... Perception p = myBody.sense(sid,myFilter1,1000); log("current value: "+p.getContent()); // observing next observable event, // which should be either // op_completed or state_changed Perception p = myBody.sense(sid,myFilter2,1000); String type = p.getType(); if (type.equals("state_changed")){ state_changed = true; } } catch (NoPerceptionException ex){ // something wrong happened in the artifact // or simply artifact too slow in executing the op... break; } catch (OperationNotAvailableException ex){ // inc was not part of artifact usage interface... break; } } ...

Fig. 7. (Left) Complete denition of the MyArtifact type; (Right) A code fragment showing an example of use of a MyArtifact artifact

the right). As declared in the artifact manual, artifacts of sfMyArtifact kind have two possible observable states, labelled as stateA and sfstateB, with the former functioning as starting state. In the stateA state, the usage interface includes both the op1 and op2 operations, while in the stateB state the usage interface includes only op2. The execution of the op1 operation causes the update of an internal counter of the artifact, whose new value is made observable by generating a new value event. When the internal counter reaches a maximum value (provided with artifact initialisation), the artifact changes its observable state from stateA to stateB. The execution of the op2 operation simply makes the current value of the internal counter observable, by generating an event of the kind value. As far as the artifact use is concerned, in the fragmentafter creating an agent body inside the working environment where the artifact is locatedop1 operation is executed repeatedly, logging each time the value perceived by observing events generated as a consequence of the operation execution, until a change of artifact state is observed. In the example, two ltersinstances of the class BasicFilter,

CArtAgO: A Framework for Prototyping Artifact-Based Environments

81

part of the utility class of CArtAgOare used to select the perceptions. Using BasicFilter, a stimulus is selected if and only if its type description matches one of the descriptions provided as parameter of BasicFilter constructor (implemented as array of strings). Run-time environment and related tools This is the part actually responsible of the life-cycle management of working environments at run-time. Conceptually, it is the virtual machine where artifacts and agent bodies are instantiated and managed that is responsible of executing operations on artifacts and collecting and routing observable events generated by artifacts. Some tools are also made available in CArtAgO for on-line inspection of working environment state, in particular artifact state and behaviour, in terms operation executed and events generated. 4.2 A Complete Example: Hello Philosophers!

To illustrate a simple but complete example of MAS application exploiting artifact-based working environments, we consider the Hello philosophers examplelisted among the basic examples in CArtAgO distribution, which is used here analogously to the (in)famous Hello world example for traditional programming languages. The example refers to the well-known problem introduced by Dijkstra in the context of concurrent programming to check the expressiveness of mechanisms and abstractions introduced to coordinate set of cooperating / competing computing agents. Briey, the problem is about a set of N philosophers (typically 5) sharing N chopsticks for eating spaghetti, sitting at a round table (so each philosopher share her left and right chopsticks with a friend philosopher on the left and one on the right). The goal of each philosopher is to live a joyful life, interleaving thinking activity, for which they actually do not need any resources, to eating activity, for which they need to take and use both the chopsticks. The goal of the overall philosophers society is to share the chopsticks fruitfully, and coordinate the access to shared resources so as to avoid forms of deadlock or starvation of individual philosopherse.g. when all philosophers have one chopstick each. The social constraint of the society is that a chopstick cannot be used simultaneously by more than one philosopher. The problem can be solved indeed in many dierent ways. By adopting the A&A perspective, it is natural to model the philosophers as cooperative agents and the tablemanaging the set of chopsticksas the coordination artifact that agents share and use to perform their (eating) activities. It is easy to encapsulate in the table artifact the enactment of the social policy that makes it possible to satisfy both mutual exclusion for the access on the individual chopsticks, and avoid deadlock situations. Fig. 8 shows the complete application, with the table artifact implemented upon CArtAgO, the agent philosophers directly implemented as at Java threads, without relying on a specic agent architecture. The usage interface of the table artifact is composed by only two operations, getChops and releaseChops, which can be used respectively to get two chopsticks

82

A. Ricci, M. Viroli, and A. Omicini

import java.util.*; import alice.cartago.*; public class Philosopher extends Thread { private int lchop, rchop; private IAgentBody myBody; private String name; public Philosopher(String name, int c0, int c1, String envName) throws Exception { this.name=name; lchop = c0; rchop = c1; ICartagoEnvironment env = Cartago.getInstance(envName); myBody = env.getAgentBody(name); } public void run() { try { ArtifactId tableId = myBody.getArtifactId("table"); SensorId sid = myBody.linkSensor(new DefaultSensor()); Op getOp = new Op("getChops",lchop, rchop); Op releaseOp = new Op("releaseChops",lchop, rchop); IPerceptionFilter myFilter = new BasicFilter("chops_acquired"); while (true){ myBody.execOp(tableId,getOp,sid); try { myBody.sense(sid,myFilter,5000); eating(); } catch(NoPerceptionException ex) { log("starved."); break; } myBody.execOp(tableId,releaseOp); thinking(); } } catch (Exception ex){ } } private void eating(){...} private void thinking() {...} private void log(String msg){...} }

import alice.cartago.*; import java.util.*; public class Table extends Artifact { private boolean[] chops; private List<PendingReq> reqs; public Table(int nchops){ chops = new boolean[nchops]; reqs = new LinkedList<PendingReq>(); for (int i = 0; i<chops.length; i++){ chops[i]=true; } } @OPERATION void getChops(int c0, c1){ if (chops[c0] && chops[c1]){ chops[c0] = chops[c1] = false; genEvent("chops_acquired"); } else { PendingReq req = new PendingReq(c0, c1, getOpId()); reqs.add(req); } } @OPERATION void releaseChops(int c0, int c1){ chops[c0] = chops[c1] = true; Iterator<PendingReq> it = reqs.listIterator(); while (it.hasNext()){ PendingReq r = it.next(); if (chops[r.c0] && chops[r.c1]){ it.remove(); chops[r.c0] = chops[r.c1] = false; try { genEvent(r.reqId,"chops_acquired"); } catch (Exception ex){} } } } private static class PendingReq { public int c0,c1; public OpId reqId; public PendingReq(int c0, int c1, OpId id){ this.c0 = c0; this.c1 = c1; reqId = id; } } }

public class HelloPhilosophers { public static void main(String[] args) throws Exception { String envName = "restaurant"; ICartagoEnvironment env = Cartago.getInstance(envName); env.createArtifact("table",Table.class,new ArtifactConfig(5)); for (int i = 0; i<5; i++){ new Philosopher("philo-"+i,i, (i+1)%nphilo,envName).start(); } } }

Fig. 8. Dining philosophers interacting within through a CArtAgO working environment called restaurant. Philosophers agents are simply implemented upon at Java threads, while the table is implemented as a table artifact of class Table.

from the table and to give them back. The inner machinery of the table artifact ensures mutual exclusion on the access on chopsticks (an artifact executes one operation at a time, analogously to monitors) and deadlock avoidance (by releasing the chopsticks only if both are available, enqueueing the pending requests). It is worth noting the way in which observable events are generated: if the chopsticks are available when an instance of getChops operation is triggered, then the event chops acquired is immediately generated, through the genEvent primitive; otherwise, the pending request is enqueued and the event is generated

CArtAgO: A Framework for Prototyping Artifact-Based Environments

83

as soon as the chopsticks become available with the execution of a releaseChops operation. On the agent side, a philosopher gets an agent body during its initialisation, and then exploits it during its main activitywhich is dened by the body of the method run. It is worth remarking that here we adopted such a simplistic implementation for agents just to make the description of CArtAgO usage and integration with agent platforms as simple and concise as possible. The same example using simpA agent framework can be found in simpA distribution. Agent main activity accounts for repeatedly alternate eating and thinking sub-activities, using the table artifact to get (and release) the chopsticks. In particular, to get the chopsticks the agent triggers the execution of the getChops operation on the table artifact, specifying a sensor (previously linked to its body) to collect stimuli related to this action. Then, it pro-actively observes the sensor for 5 seconds using the sense primitive, ltering stimuli that concern chops acquired events. If no perception is sensed within 5 seconds, the philosopher starves and terminates. Otherwise, it performs its eating activity and then, after eating, it releases the chopsticks by executing a releaseChops operation on the table artifact, without specifying any sensor (since, in this simple implementation, it is not interested to observe the eects of such an action).

Related Works

The approach based on artifacts shares the same engineering aims introduced by Weyns and colleagues in [11], where they identify a general model and an architecture that can be (re-)used to engineer environments in MAS, despite of the specic application domain. The model presented by the authors is concernbased : the environment is modelled as a set of modules that represent dierent functional concerns of the environment. A similar focus, but in some sense less general, can be found also in the work of Platon and colleagues [12], where a general model for environments providing functionalities for over-hearing and over-sensing is presented. Our notion of artifact could be compared at a rst glance with the notion of functional modules describe by Weyns and colleagues. The main dierence is that artifacts are conceived to be rst-class abstractions both for the engineers designing and programming agent environment and for the agents using such an environment : agents do not perceive the environment as a single entity providing a set of functionalities (which are internally engineered upon a set of modules), but directly create, share, use, manipulate, destroy artifacts, each designed to encapsulate some kind of function. The model for perception and sensing described in the paper shares many points with the modelmore generaldiscussed in [9], introducing the notion of active perceptions. Such a model decomposes perceptions into a succession of three functionalities: sensing, interpreting and ltering. First, sensing maps the state of the environment to a representation. The agent can select a set of foci, that enable the agent to sense specic type of data in the environment. The representation of the state is composed according to a set of perception laws, that

84

A. Ricci, M. Viroli, and A. Omicini

can be used by designers to enforce specic constraints on perceptions. Then, agents interpret representations by means of descriptions, that are blueprints that map representation onto percepts, modelled as expressions that can be understood by the internal machinery of the agent. Finally, agents can select a set of lters, to restrict the perceived data according to specic context relevant selection criteria. In our model, sensors and sense actions provide some of the functionalities discussed above. In particular, by following the meaning introduced by the authors, each sensor can be used as a specic focus: the idea is that an agent can dynamically create and link to their body dierent kind of sensors, with distinct features (such as buering, ltering, etc), to partition the perceptions from the environment, in our case related to artifacts (even if the model can be extended to consider also perceptions directly related to other agents). Similar to perceptual laws discussed above, sensor activity can be constrained according to laws enforced by the organisational and physical context where the agent is situated: this aspect will be explored in future work, with the introduction of an explicit support for organisation modelling on top of workspaces (see Sect. 6). Pattern-driven sensing described in Sect. 3 could be framed as a simplied form of ltering as dened in active perceptions, with some points that concern also interpretation: patterns act as simple lters that agents can specify to fetch in a data-driven way the data collected by sensors, and require that an explicit description is adopted for describing the events or stimuli posted to sensors. Finally, the artifact abstraction and CArtAgO framework draw on the research work on tuple centres as programmable tuple-based coordination media and TuCSoN coordination model [13]. Artifacts can be framed as a generalisation of the notion of tuple centre: more precisely, tuple centres can be conceived as a type of coordination artifacts [8], as artifacts designed to encapsulate programmable coordination services.

Concluding Remarks

In this paper we described CArtAgO, as a framework supporting the engineering of artifact-based working environment in MAS. First we described the abstract model and architecture of the framework, and then a rst basic prototype technology implementing most of the core functionalities. Among the issues not considered for lack of spaceand that can be found in the artifact conceptual frameworkwe mention here: (i) artifact composition support for linking together existing artifacts to dynamically compose complex artifacts, by dening and exploiting artifact link interfaces ; (ii) artifact management support for inspecting, controlling, testing artifact state and behaviour, by dening and exploiting artifacts management interface, besides usage interface. Among the issues not currently faced in CArtAgO, and that will be part of our future work, we mention here distribution, security, and organisation. As far as distribution is concerned, currently in CArtAgO there is no explicit account for the way in which workspacesand possibly also artifactscan be

CArtAgO: A Framework for Prototyping Artifact-Based Environments

85

distributed across multiple nodes of a networks. In current version, working environments are conned to a single CArtAgO virtual machine (node) and then agents can create and use artifacts that live in their local environment. Distribution is achieved by ad-hoc linking of artifacts, that is, by exploiting network connections that makes it possible the low-level communication among artifacts belonging to dierent workspaces and working environments. In future work we will focus on extending the framework towards a full-edged infrastructure, providing a rst-class support for such an aspect. Security and organisation are related issues, and call for the introduction of an explicit support for organisation built on top of the basic CArtAgO abstractions, which would be eective also to model security aspects such as access control. By drawing on our previous research work about such aspects on TuCSoN infrastructure, in CArtAgO we plan to introduce a role -based model, inspired to RBAC (Role-Based Access Control) architectures [14], such as RBAC-MAS [15]. Such a model will be based on the notion of workplace. A workplace denes the set of roles and related organisational rules or contracts being in force in a workspace. The contracts denes, in particular, the norms and policies that rule agent access to the artifacts belonging to the workspace. For example, depending on the role(s) that an agent is playing inside the workplace, it may have or not the permission to use some artifacts or to execute some specic operations on some specic artifacts. So, workplaces would dene an organisational layerand, consequently, a security layeron top of workspaces. Besides the above three main points, future work will be devoted also: (i) to improve the development of the prototype, including all the missing features that are currently part of the abstract modelsuch as the support for workspaces and future extensionssuch as workplaces; (ii) to integrate existing services as kinds of artifact, in order to be easily re-used when engineering applications on top of CArtAgO: an example is given by artifacts wrapping TuCSoN tuple centres, providing agent coordination facilities; (iii) to dene suitable formal models and ontology for describing function descriptions, operating instructions, and observable state description, possibly reusing existing research eorts on service description models and (standard) languages, such as OWL-S. Finally, existing and ongoing research in environment for MAS will be important to improve the theoretical foundation of CArtAgO, concerning the notion of artifact and related concepts: for instance, the research work on active perceptions can be important to improve and extend the model of sensing currently adopted.

References
1. Ricci, A., Viroli, M., Omicini, A.: Programming MAS with artifacts. In Bordini, R.P., Dastani, M., Dix, J., El Fallah Seghrouchni, A., eds.: 3rd International Workshop Programming Multi-Agent Systems (PROMAS 2005), AAMAS 2005, Utrecht, The Netherlands (2005) 163178 2. Weyns, D., Omicini, A., Odell, J.: Environment as a rst-class abstraction in multi-agent systems. Autonomous Agents and Multi-Agent Systems 14 (2007) 4960 Special Issue on Environments for Multi-agent Systems.

86

A. Ricci, M. Viroli, and A. Omicini

3. Viroli, M., Ricci, A., Holvoet, T., Shelfthout, K., Zambonelli, F.: Infrastructures for the environment of multiagent systems. Autonomous Agents and Multi-Agent Systems 14 (2007) 530 Special Issue on Environments for Multi-agent Systems. 4. Nardi, B.A.: Context and Consciousness: Activity Theory and Human-Computer Interaction. MIT Press (1996) 5. Kirsh, D.: Distributed cognition, coordination and environment design. In: European conference on Cognitive Science. (1999) 111 6. Agre, P.: Computational research on interaction and agency. Articial Intelligence 72 (1995) 152 7. Amant, R.S., Wood, A.B.: Tool use for autonomous agents. In Veloso, M.M., Kambhampati, S., eds.: AAAI/IAAI05 Conference, Pittsburgh, PA, USA, AAAI Press / The MIT Press (2005) 184189 8. Omicini, A., Ricci, A., Viroli, M., Castelfranchi, C., Tummolini, L.: Coordination artifacts: Environment-based coordination for intelligent agents. In: AAMAS04. Volume 1., New York, USA, ACM (2004) 286293 9. Weyns, D., Steegmans, E., Holvoet, T.: Towards active perception in situated multiagent systems. Applied Articial Intelligence 18 (2004) 867883 10. Viroli, M., Ricci, A.: Instructions-based semantics of agent mediated interaction. In: AAMAS04. Volume 1., New York, USA, ACM (2004) 286293 11. Weyns, D., Holvoet, T.: Formal model for situated multiagent systems. Fundamenta Informaticae 63 (2004) 125158 12. Platon, E., Honiden, S., Sabouret, N.: Oversensing with a softbody in the environment: Another dimension of observation. In Kaminka, G.A., Pynadath, D.V., Geib, C.W., eds.: Workshop Modeling Others from Observation (MOO 2005), IJCAI-05, Edinburgh, Scotland (2005) 13. Omicini, A., Zambonelli, F.: Coordination for Internet application development. Autonomous Agents and Multi-Agent Systems 2 (1999) 251269 14. Sandhu, R., Coyne, E.J., Feinstein, H.L., Youman, C.E.: Role-based control models. IEEE Computer 29 (1996) 3847 15. Omicini, A., Ricci, A., Viroli, M.: An algebraic approach for modelling organisation, roles and contexts in MAS. Applicable Algebra in Engineering, Communication and Computing 16 (2005) 151178 Special Issue: Process Algebras and Multi-Agent Systems.