Intelligent Interactive Technology

Multi-Agent  Platform  based  on  Deep  Behaviour  Projection  (DBP-MAP)

Our principal target is a systematic description of increasing levels of agent reasoning abilities, where a given behaviour can "grow" into an expanded level, while every new level can be projected onto a deeper (more basic) behaviour. More precisely, we develop a scalable Deep Behaviour Projection (DBP) agent architecture, and enhance agents' behaviour incrementally. In particular, we attempt to investigate conditions for a rigorous comparative analysis between deliberate and reactive agents.

The DBP framework has the following properties:

• realisation of the behavioural programming principle (predictably translating symbolic descriptions into the low-level controllers specifications or reactive behaviours);
• encapsulation of an emergent behaviour (a feedback between the emergent behaviour and symbolic descriptions);
• behaviour depth or duplication (both embedded and emergent levels), unlike ``shallow'' behaviour subsumption;
• behaviour projection and resultant functional interchangeability, mediated internally within the agent.

The feedback connection from an emergent behaviour to a (meta-)action theory, and then to a derived behaviour on a higher level, extends the behavioural programming principle. Thus, a successful agent behaviour can be present in the architecture in two forms: implicit (emergent) and explicit (embedded). We believe that this duplication (or depth) provides necessary functional interchangeability, and allows the agent to "mediate" among related behaviours. 

The following list illustrates the hierarchy with informal examples highlighting some obvious biological parallels:

• tropistic behaviour: Sensors => Effectors (plants following sunlight, spiders catching prey);
• hysteretic behaviour: Sensors & Memory =>  Effectors (movement of a school of fish, ants tracing pheromones);
• task-oriented behaviour: Sensors & Memory & Task =>  Effectors (lions hunting, patrolling territory);
• domain-oriented memory projection: Sensors & Memory & Domain => Memory, followed by behaviour-based actions (whales communication during a hunt)

The Deep Behaviour Projection framework underlies hierarchies like this and ensures that more advanced levels capture relevant behaviour more concisely than their deeper projections. In short, an agent's complexity is relative: for any elaborate agent type, it is possible to define more concisely another agent type with the same external behaviour. In practice, an agent based on the DBP architecture will always have a choice as to which one of related hierarchical levels should assume control to better suit external environment. This choice serves, in fact, as the "middle-out" layer between any two levels of the DBP agent architecture --- replacing the need for a generic hub connecting "reactive behaviour" and "cognitive skills".

The DBP framework was used in design and development of Cyberoos and the SATP architecture. Watch this space for more details on our generic DBP-MAP technology developed for Windows-NT.