Dimensions in programming multi-agent systems

J. F. H. and R. H. B. would like to thank CNPq and CAPES for partially funding their research.

We invite the reader to refer to the proceedings of the EMAS Workshop series and its predecessors for a broad overview of the area.

The acronym JaCaMo comes from combining the names of the three platforms on which JaCaMo is based, namely Jason, Cartago, Moise.

In fact, it is also physically a locality, since a workspace runs on a single host.

More details about the concepts in the agent dimension can be found in Bratman (1987), Rao (1996), Labrou et al. (1999), Bordini et al. (2007).

More details about the concepts in the environment dimension can be found in Weyns et al. (2007), Omicini et al. (2008), Ricci et al. (2009).

More details about the concepts in the organization dimension can be found at Demazeau and Rocha Costa (1996), Lemaître and Excelente (1998), Hubner et al. (2002), Boissier et al. (2007), Rocha Costa and Dimuro (2009).

This assumption can be relaxed by defining an ‘interaction’ language (not only for agent communication but also for perception and action), some sort of common language or application ontology that defines a set of goals, actions, and percepts that are translated into environment observable properties and operations, into agent beliefs and actions, and into agent and organization goals. It can also be relaxed if agents are capable of learning, possibly from the application ontology, what g34 means. For the particular example of plans for required social goals, it is worth mentioning that in MAOP agents often acquire plans at runtime through interaction with other agents or a planner (e.g. encapsulated as an artefact). This kind of sophistication, however, is not considered in this paper and we assume the developer uses matching identifiers to link the dimensions.

The code for the examples used in this paper is available at http://jacamo.sf.net/ker2017. The technical documentation of the JaCaMo platform with all the programming constructs that can be used to write a MAOP is available on the JaCaMo website at http://jacamo.sf.net/

The syntax of AgentSpeak, the language that inspired our agent language, was in turn inspired by Prolog and thus identifiers starting in upper case are variables.

Note that this can also be done in an agent program as shown in the previous section with a particular language construct called internal actions. We do not show it here due to space constraints.

If an agent requests an operation which is provided by multiple artefacts (located in the workspaces that the agent has joined) and without explicitly stating which one is meant, no error occurs and one is selected non-deterministically by the infrastructure.

This assumes full cooperation. However, also different scenarios can be handled, where, for example, an agent can reject a goal delegated by another agent, or one which the organization forbids.

We use the .broadcast command instead of .send, which is used to send a message to all agents in the system.

In the agent code, t1 is the name of the artefact, whilst ArtId is the identifier of the artefact whose name is t1. The value of ArtId is determined by the underlying infrastructure that manages artefacts.

Of course we could change the protocol in the message-based version too, but it would require a lot of extra message passing that are avoided by the features provided by artefacts.

JaCaMo provides a library of plans that enable the agents to react to the state of the norms managed by the organization. For example, this library has a plan that creates goal g for the agent whenever the agent is obliged by the organization to achieve g. The use of this library is optional, as in other applications agents may prefer to decide weather to adopt the goal or not based on their own reasons.

For space reasons, we do not explain how agents adapt and redefine the specification of their organization, although this is possible and can be used for interesting forms of self-organizing systems.

For a overview of the relevant literature, see Bordini et al. (2005), Bordini et al. (2006), Dastani (2015).

Some of the concepts used in the modelling are available in some platforms (e.g. teams in JACK and norms in 2OPL). However, the set of concepts in these platforms is limited when compared to those used in modelling (e.g. role and protocols are not available in most of those platforms).