Draft

Definition of basic agent capability list

Agents, as used in this context, are high-level software entities with defined roles. They interact both with the physical world and each other. We include the humans that are part of this physical world.

Separate technologies have been developed for computing machines to deal with humans and with mechanical machines. Fig. 1 below describes how the FIPA work in the agent field can be visualised.

Fig. 1 - Agent and society of agents

The London meeting (18-19 April) produced a list of 16 basic agent capabilities that participants believed could be the target of a short-term (December 1997) standardisation. Such an assessment did not obviously imply, at that point in time, any judgement on the desirability of nor commitment to standardising any of those items. Steps in that direction will be taken at the next meeting (Yorktown, NY, 24-26 June).

Below is an attempt at defining the 16 items and providing, where helpful, some examples.

1. speech processing primitives (in & out)

Because intelligent agents communicate with humans and speech technology is well advanced therefore the interface with the agent is an issue for standardisation. Examples of primitives are those related with the linguistic, emotional and referential content.

1. visual input primitives

Visual input encompasses a large number of input possibilities. Substantial technological progress has been achieved. The areas included:

• identification of pre-specified objects and their position
• hand-writing and graphics recognition
• face, gesture and expression recognition

1. visual output primitives

The essence of good human-computer interaction design is to manage the cognitive complexity of the information that is visually conveyed to the user. Many examples can be stated, such as

• facial animation
• graphical object animation
• use of the colour dimension on the screen

1. active (or controllable) head-eye system

Visual input is essential for a number of applications. To be sufficiently flexible, the visual input must be controllable from other parts of a system to provide the desired information. Among functions and variables to specify are: control for stereo cameras of version and vergence, separate control of eyes (cameras) and head (combination), ability to perform saccades, fixation and smooth pursuit. These and other control modes allow the system to actively browse the environment, search for objects, make detailed analyses and respond to requests by the system.

1. actuator primitives

This is needed to affect the physical environment. The range of possibilities include grasping and lifting to physical movement.

1. fusion of information

We do not expect to deal with only one type of input information at a time. Multiple sources of inputs have to be correlated with each other. Fusion can take rather simple forms such as synchronisation of events to complex algorithmic multi-sensor data fusion.

1. support of cultural diversity

When agents deal with the human part of the physical world then the issue of style becomes important. For example even within a single culture there are different styles when dealing with children or adults. Indeed different language support provide a natural cultural dimension.

1. finite set of basic message types/communication primitives for inter-agent communication

Formalised approach for the interchange of knowledge between agent. For example a language based on speech act vocabulary or KIF.

At a lower level of abstraction for interchanging information between agents, various performatives are normally used. KQML is an example of such a set of performatives.

1. composition operators for basic message types

This refers to the composition of basic message types into protocols for agents to cooperate with one another e.g. agents negotiating with each other for optimum use of resources.

1. standard library of interaction protocols

Multi-agent systems will be used in a variety of applications that are capable of being classified into two specific types. Standard libraries will help application developers with the necessary implementation tools.

1. refinement of interaction protocols for specific tasks

Standard libraries can provide subsets of protocols to support typical applications. Examples are supplied by Web filters, a diary assistant, space planning etc.

1. semantic information representation and cognitive architectures

Agents will need to represent a large set of agent-linguistic primitives and to make use of this knowledge in scheduling their own task. This representation has some degree of implication on the internal structure of the agent.

1. reference programming libraries

These will allow users to speed up development of specific agent applications.

1. agent programming human interface

Agent development systems are an important tool for implementing applications. An explicit standardisation may not be strictly necessary but beneficial in the early phases of the industry.

1. specification of minimal class library

Again a specification of a minimal class can benefit the early phases of industry.

1. aspects of society of agents

When large number of agents interact with one another then some of the following issues may generate their own scope for standardisation

• authentication, privacy, security, safety
• authority and delegation
• etiquette and personality (e.g. emotional states)

Back to FIPA homepage

Dr. Leonardo Chiariglione		+39 11 228 6129/6116/5111
Multimedia and Video Services		+39 11 228 6299/6190/5520
CSELT		leonardo.chiariglione@cselt.it
Via G. Reiss Romoli, 274		http://www.cselt.it/ufv/leonardo
I-10148 Torino (ITALY)