Go back to: About Us > Departments > Academic Departments > School of Creative Technologies (CT) > Research

School of Creative Technologies (CT)

Intelligent Agents Group

Introduction

In contrast to many approaches towards agent construction, in which the 'agents' seem little removed from software daemons, work in the Intelligent Agent Group has to date concentrated on the production of 'psychologically plausible' agents. Such agents have been deployed in place of human players within large scale simulations (e.g. ModSAF and JSAF), as well as implementing decision support systems. The ultimate level of modelling is the replacement of 'man' in situated, 'man-in-the-loop' control systems.

The agents we have built to date are implemented as sets of rules, executed within the Soar production system architecture to provide real-time behaviours. Initially, work concentrated on models of single individuals (for example, a shipboard anti-aircraft warfare officer (AAWO) performing certain decision and communication tasks during a simulated air battle). More recently, the focus has turned towards reusable models of teamwork within Soar which may be used to underwrite communication and coordination protocols within teams of interdependent agents in a domain independent way.

Group Members and Contacts

Tony Kalus | Dan Allsopp | Nipan Maniar

Areas of Interest

The following list describes some of the key aspects of our work:

Agent based simulation

of particular interest to us are simulations of the real world, populated by entities that correspond to people, vehicles etc. Such simulations may be used for modelling crowd behaviour, air or road traffic, land/sea/air combat, building or aircraft evacuation etc. Our work involved modelling the behaviour of the human entities within the simulated environment in a cognitively realistic way. Such approaches are becoming increasingly popular with the advent of 'simulation based acquisition' programs. They also provide a powerful way of testing the effects of changes in agent strategy and tactics, as well as changes to the environment. As a training tool, synthetic environments allow trainees to try out their own decision-making procedures within the context of a population of 'human-like' entities who may be expected to react in plausible and appropriate ways.

Knowledge acquisition for agent construction

The agents we are interested in constructing are knowledge-based systems. Extracting appropriate knowledge from subject matter experts and other primary sources (frequently in the form of documentation) is traditionally an arcane process. One possible line of approach is to identify the required functionality of an agent with an individual who has received effective training in only the necessary and sufficient skills required for that role. This approach has several spin-offs for the more general management of knowledge within an arbitrary organisation.

Specification and storage of task knowledge for agent construction

Once task knowledge has been extracted from subject matter experts and other documental sources, it is useful to specify this knowledge in an agent language independent representation to facilitate reuse. This way, there should be no need to return to source documents again and again once task knowledge has been captured correctly. We have been using the CommonKADS methodology to specify agent tasks, and have devised a preliminary database architecture to store these specifications. Agent code in a range of languages can be and have been derived from specifications stored in this task repository.

Formal models of coordination, command and control

Ensuring appropriate and effective communication within multi-agent systems is approached through formalised notions of teamwork. Once the infrastructure for a generic team has been established, the rapid creation of domain based teams becomes possible. Aligned with this work is the identification of more general command and control protocols for a given organisation (or type of organisation). Once defined, these protocols may be implemented once only for a given implementation language and reused as required for any similarly organised multi-agent system within that language.

Group Presentations & Publications

Tony Kalus & Tony Hirst

Soar Agents for OOTW Mission Simulation, 4th International Command and Control Research and Technology Symposium, Näsby Slott, Sweden. September 1998.