Designing Agents and Agent Systems

All Messages

How important are architectures
Joerg Mueller -- Tuesday, 10 February 1998, at 9:27 a.m.

How important do you feel is the architecture of an agent compared to the underlying theory, the knowledge representation formalism, or the problem-solving functionality (domain function) of the agent? Do you think that there should be a relationship (that can be exploited during design time) between the architecture of an agent and the types of application in which an agent is to be used?

Re: How important are architectures
Rodrigo Ventura -- Monday, 23 February 1998, at 7:55 p.m.

The purpose of an agent theory is to provide a conceptual framework general enought to handle any problem. This framework shall not be domain dependent. Que question is how far is it possible to advance with this framework without narrowing the range of aplications.

Re: How important are architectures
H.J. Mueller -- Tuesday, 24 March 1998, at 9:17 p.m.

There are people who say, that architectures are nothing but angel dust. All that counts are formal specifications. ......Well, I say that there is a need for a guideline if you try to build up the specification. Why not using an architecture?

Counter question: Why is the IMAGINE architecture (Mouth, Head, Body) or the ARCHON architecture so successful? Because they separate the pure agent functionality and the domain stuff. And they are open to integrate existing systems. Hence, the developer will get a simple framework with a set of basic functions and a docking station for domain specific funcionalities which he already has.....By the way, that's the same idea as in CommonKADS for knowledge based systems. (and there are approaches which extend the CommonKADS model set for MAS development!)

Re: How important are architectures
Fabrice Chantemargue -- Wednesday, 25 March 1998, at 4:44 p.m.

I think the architecture of an agent is very important because it can inform on the flexibility of the system. It can give information on the scalability of the system and on its robustness (fault tolerance).

May be, we should even consider the organization of the agent prior to its architecture. For instance in the case of physically embodied agents (robots), the organization of the agent would stipulate an agent is made of a body (encompassing various modules such as a control algorithm, some kind of internal state, a perception module ...) and some effectors and sensors to act and perceive in its environment. Then, the architecture of a particular agent belonging to this class of organization would specify the types of sensors and effectors, e.g. whether the agent has a bumper, an Infra-Red sensor, a gripper ...

If there is no relationship between the organization (as defined here) of an agent and the types of applications in which this agent is to be used, I think it is clear that there is a relationship between the architecture and the applications in which this agent is supposed to be involved.

---

individual vs. social
Cristiano Castelfranchi -- Thursday, 19 March 1998, at 4:40 p.m.

How to reconcile individual autonomy and self-interest with collective interest and performance?

Re: individual vs. social
Rodrigo Ventura -- Thursday, 19 March 1998, at 7:47 p.m.

I guess that there cannot be an a priori answer to that question. It is not only strongly domain dependent, but also depends on the agent's behavior. I think that the best way to tackle that problem is to allow the agents the awareness of their own individual/collective behavior. And this awareness must result of some sort of evaluation of the results of the collective. In other words, an agent shall not have a fixed level of "cooperativeness". It must be dynamical in this sense.

Can such a mechanism be called a meta-level reasoning about its own behavior?

Re: individual vs. social
Fabrice Chantemargue -- Wednesday, 25 March 1998, at 5:55 p.m.

I think it depends on the type of autonomy of the agents. In the case of operational autonomy (the agent is free to choose its action at any time among actions provided by the designer), I think the collective performance is not incompatible with the individual autonomy of the agents. Several experiments have indeed shown how collective phenomena (cooperation for instance) emerge through the recurrent actions of the agents in the environment (some kind of self-organizing systems). As the designer specifies the set of possible actions, an "adequate tuning" is possible according to the application. In the case of behavioral autonomy (the agent is free to choose its actions at any time and moreover it should have been able to build its actions at least partly), it is much more complicated since the designer has no control (or very weak control) on the set of possible actions the agent can take.

Regarding a kind of evaluation by every agent on what results from the collective seems to me very difficult to be set up. For instance, when an agent should decide to analyse the results of the collective? How can you ensure the period of observation (analysis) is relevant?

---

Maturity of Agent Design?
Michael Luck -- Thursday, 19 March 1998, at 7:09 p.m.

How mature is the area of agent design? Do we yet have a critical mass of stock designs that allow systems to be constructed from off-the-shelf components, or must we start from scratch for each application or domain?

Does it matter if more and more agents are designed, with only slight differences in functionality? Might this ultimately lead to the fragmentation of the field?

Re: Maturity of Agent Design?
Fabrice Chantemargue -- Wednesday, 25 March 1998, at 6:10 p.m.

Unfortunately, my feeling on this point is that we have to start from scratch the agent design for almost every application. I am wondering whether we will experience the same things as in OO techniques: Who has ever experienced the reusability of code so many times put forward in OO techniques?

---

Correlations
H.J. Mueller -- Tuesday, 24 March 1998, at 8:49 p.m.

What are the attributes of application problems, that force a MAS realization? .....Other way around Which characteristics of agents and MAS would be necessary to solve a specific application specified in some requirement description language? How deep must be such specifications?

In summary: What are the correlations between application problems and agent (MAS) attributes?