What is an intelligent agent? This apparently anodyne question provokes a great deal of debate on and off the Internet. While simplifying to the maximum, we find on one side the people that almost regard intelligent agents as human beings, and on the other those that consider them to be no more than simple software programs.
Jacques Ferber proposes the following definition: " Something can be called an agent if it is a physical or virtual entity that:
- is able to act in an environment,
- can communicate directly with other agents,
- is driven by a group of tendencies [... ],
- has its own resources,
- is able to perceive its environment (but in a limited way),
- has only a partial representation of this environment (and possibly none),
- has skills and offers services,
- may be able to reproduce,
- whose behavior aims to satisfy its objectives, by taking account of the resources and skills that it possesses, and according to its perception, of its representations and the communications that it receives."
On the other hand, many regard an agent as an "entity authorized to act in the name of someone else". Such a definition puts an intelligent agent in the same category as a real estate agent, a security agent, or a sales agent... Consequently, the distinction between an intelligent agent, and basic software remains very unclear.
In spite of its limits, this vision constitutes a good starting point for a definition that would be sufficiently realistic, without being too limited. We can therefore affirm that an intelligent agent is a software entity that has its own attributes, and that acts with the goal of performing a certain number of tasks in the name of another entity (another agent or a person).
- Autonomy: an agent is able to take initiative in an autonomous way and it exerts control over its actions. In other words, it has the capacity to define its own objectives and achieve them without human supervision.
- Capable of collaboration: an agent is able to interact and cooperate with other agents or human beings in order to complete its tasks.
- Capable of learning: an agent is able to adapt to the needs of its user by analyzing its past actions.
- Goal-oriented: an agent accepts precise "human requests" and decides how to satisfy them.
- Flexibility: the actions of an agent are not entirely pre-established and defined; it is able to choose what it will do, and in which order, according to the external environment.
- Self-starting: contrary to traditional software, an agent can decide to start an action at a precise moment, according to the external environment,
- Individual character: an agent has a well-defined personality and its own "emotional state",
- Mobility: an agent is able to move from one system to another and through various architectures and platforms, etc.
But when should we speak about intelligent agents? Do they have to have all, or just some, of these attributes?
The debate is interminable, perhaps insoluble, and not of much interest to the end-user. Many companies, for obvious marketing reasons, make use of this terminological blur to qualify their software as an intelligent agent. What is clear is that, as of this date, on the Internet, no agent known as "intelligent" has all of these attributes. Consequently, trying to define an "intelligent" agent creates an oxymoron. Between the theoretical principles of artificial intelligence and concrete reality that one observes on the Internet, the difference can be great.
That said, as Pierre-Alain Le Cheviller points out, "these agents, if they don't use the same technology [as artificial intelligence], have the same objectives and can be very useful tools for information professionals".
For this reason, instead of speaking about intelligent agents in a broad sense, it is undoubtedly preferable to speak about software agents or better still to refer to the tasks that these assistants are intended to perform: therefore we will talk about search agents, monitoring agents, e-commerce agents, etc.
An extract from "Strategic Internet Intelligence" (Dunod, 2000)
With the author's agreement.