An Emotional Agent the Detection and Control of Emergent States in Autonomous Resource{bounded Agents Thesis Proposal Research Progress Report 3

In dynamic and unpredictable domains, such as the real world, agents are continually faced with new requirements and constraints on the quality and types of solutions they produce. Any agent design will always be limited in some way. Such considerations highlight the need for self-referential mechanisms, i.e. agents with the ability to examine and reason about their internal processes in order to improve and control their own functioning. Some emotional states, such as debilitating grief, could be considered as disruptive responses to new circumstances. In other words, the design is unable to `cope' adaptively. This research wishes to investigate the generation of such `emotional' states, their detection by self-referential mechanisms, and possible coping, prevention or solution strategies within a unifying framework of an extant broad agent architecture. There are many theories of emotionality. Sloman's (1981) `attention lter penetration' theory provides an architectural framework for the understanding of `emotional' states that involve partial loss of control of attention { so-called perturbant states. Beaudoin (1994) has designed an agent architecture (NML1) that, when implemented, should exhibit such states. Smith (1986) states three requirements for self-referential systems: a theory of the self, the use of this theory in guiding the behaviour of the system, and the ability to swap between action in the world and reasoning about the self. The NML1 design only meets the nal requirement. Also, the design requirements for self-control of problematic processing states have yet to be systematically addressed. Therefore, this work will implement a prototype agent architecture, exhibiting some kind of perturbant state, based on a revised NML1 design. The design will be extended to meet the requirements for self-referential systems. The aim is to design and partially implement an NML1 that is self-referential, can exhibit `emotional' states, detect such states, and try to do something about them. Results from this research will contribute to autonomous agent design, emotionality, internal perception and meta-level control; in particular, it is hoped that we will i. provide a (partial) implementation of Sloman's theory of perturbances within the NML1 design, ii. investigate the requirements for the self-detection and control of processing states, and iii. demonstrate the adaptiveness of, the need for, and consequences of, self-control mechanisms that meet the requirements for self-referential systems. 2