THE IMPACT OF DURATIVE VARIABLE STATE ON THE DESIGN AND CONT ROL OF ACTION SELECTION - AIBACS

1.We created psychologically-realistic artificial models of the onset, persistence and decay of emotion and drive response levels. We called this the Dynamic Emotion Representation (DER). These models are available as source code from our web page, and described in a number of publications so that they can be reimplemented for other systems. The basic representations a general-purpose framework for drives and emotions. We also provide XML scripts of events and emotions systems, and parameter settings which demonstrate the variety of behaviours and expressions generated by a face avatar. Behaviour varies appropriately depending on the avatar's prior emotional state and/or experiences.

2.We integrated a simplified version of the emotion and drive model into a current state-of-the-art dynamic-plan-based action-selection system, which is also available for download. We have demonstrated improved efficiency in meeting a variety of conflicting goals in a dynamic environment as a consequence of creating a simple network of durative state modules --- one regulating each goal. AI action selection is improved by providing these emotion-inspired pieces of decision-state memory, which are both more persistent than that found in even the best dynamic plans, but more responsive and contextually-appropriate than decisions based on ordinary memory (e.g. conventional knowledge bases or belief systems).

3.We produced a graphical system for humanoid Virtual Reality (VR) facial characters and personality expression which incorporates and depends on the full DER. We have conducted and published tests of the believability of the emotional models on ordinary human subjects. Although we began with a basic framework for the VR system, the work here went beyond our expected goals. In addition to incorporating the emotion/drive system into the humanoid and creating both GUI and XML scripting tools for controlling it (e.g. parameter setting, initiation and debugging), we also conducted novel graphics research for blending the facial expressions.

4.We have produced agent-based simulations of primate social interactions which tests the evolutionary efficacy of such representation systems. This is ongoing work, but has been presented at several primatology meetings as well as social simulation meetings. In addition, we have performed more basic research on the efficiency of this system, which has also been published and is being submitted further. We believe that new emotional states are required to regulate new evolutionary goals. For example, social emotions have evolved to regulate the social behavior of social species. Different drives/emotions help social agents avoid the dangers of both conflict and isolation. On the other hand, drives for sex and prestige encourage individuals to engage in individually-dangerous behavior that benefits their progeny and therefore their species or genes.

5.We have also put considerable effort into making the building of such complex models in agent-based simulations easier, both by creating a new research platform on top of an existing system (BOD / MASON) and by creating an IDE for dynamic plans. This work is downloaded regularly and has been used in applications as varied as computer game characters and military logistics simulations.