WP1: Architectures for self reflection and self motivation

This workpackage has contributed representations of self-knowledge, of actions that alter this knowledge, and of desires for self-extension and task-based action. These representations will be integrated into a computational account of motivation for a robot, produced with reference to a system architecture. Building on the work of [1], our account included the following features: beliefs about the system and its world, both specific and general, and beliefs about beliefs, in various forms and distributed over various subsystems (i.e. beyond a purely logical database of beliefs); a small set of intrinsic desires that specify the general types of behaviours the system can engage in; a collection of semantically rich structures representing instantiations of the intrinsic desires from various subsystems, called desire instances; a set of concurrently, asynchronously active processes that generate, activate or reactivate desire instances based on intrinsic desires and the current context, called desire instance generators; a set of desire management processes that filter, inspect, assess and select desire instances for subsequent processing (including action); and a set of intentions which represent a set of selected desire instances that the system intends to act on. Although we are using the terminology of BDI systems, our heterogeneous architecture-based approach should be distinguished from the purely logic-based approaches of such systems.

Given the system-wide aims of this workpackage, it has interacted with aspects of all of the objectives of this project, but it has been particularly concerned with the following objectives. To contribute towards a unified framework for belief representation we will develop a unified framework for representing a system’s intrinsic desires and desire instances where these can refer to both beliefs and beliefs about beliefs. To contribute towards a thoery of trading off curiosity and task demandsan architectural framework for the generation of desire instances from either task-driven or curiosity-driven intrinsic desires, along with approaches for managing competing and conflicting desires. Many more intrinsic desires could be considered in more general work, such as physical well-being (homoeostasis) or social pressures. To contribute towards actions that alter epistemic states we investigated how actions that stem from the adoption of desire instances into intentions alter the belief state of the system and other agents (including beliefs about beliefs). We evaluated our work using implementations of our approaches. In the project’s implemented systems the desire instance generator and management mechanisms provide architecture-wide control. This has contributed towards the project’s practical aims for its demonstrators.

[1] I. Wright, A. Sloman, and L. Beaudoin. Towards a design-based analysis of emotional
episodes. Philosophy Psychiatry and Psychology, 3(2):101–126, 1996. Repr. in R.L.Chrisley
(Ed.), Artificial Intel ligence: Critical Concepts in Cognitive Science, Vol IV, Routledge, Lon-
don, 2000.