The `Radical Agent Concept' in this chapter is that communication between agents in a MAS should be the simplest part of the system. When extensive real-time coordination between modules is required, then those modules should probably be considered elements of a single modular agent rather than as agents themselves. The advantage of this distinction is that system developers can then leverage standard software-engineering practices and more centralized coordination mechanisms to reduce the over-all complexity of the system. In this chapter I provide arguments for this point and also examples, both from nature and from my own research in building modular agents.

Explanations of cognitive processes provided by traditional artificial intelligence were based on the notion of the knowledge level. This perspective has been challenged by new AI that proposes an approach based on embodied systems that interact with the real world. We demonstrate that these two views can be unified. Our argument is based on the assumption that knowledge level explanations can be defined in the context of Bayesian theory while the goals of new AI are captured by using a well established robot based model of learning and problem solving, called Distributed Adaptive Control (DAC). In our analysis we consider random foraging and we prove that minor modifications of the DAC architecture renders a model that is equivalent to a Bayesian analysis of this task. Subsequently, we compare this enhanced, "rational", model to its, "non-rational", predecessor and a further control condition using both simulated and real robots, in a variety of environments. Our results show that the changes made to the DAC architecture, in order to unify the perspectives of old and new AI, also lead to a significant improvement in random foraging.

How can we build more robust reasoning systems? One approach is to build systems as layered agent architectures where each layer observes the activities occurring in the layers beneath, and uses knowledge about how to interpret these observations to control and debug those activities. Architectures with such 'towers of reflection' are uesd because it is often difficult to assure perfect operation in any one layer, and therefore an additional layer that reflects on it can be added to help cope with bugs within lower layers. In this paper we consider three layer architectures, ones with reactive, deliberative, and reflective layers. In the deliberative layer, any particular method of inference, strategy for producing a plan, or other method of deliberation is likely to break now and then. We propose that reflective critics are a primary type of agent in the reflective layer that is concerned with noticing problems in the recent activities within the deliberative layer. In this paper we explore what types of reflective critics there are and how they can be represented declaratively. We illustrate the use of these reflective critics with a set of scenarios drawn from problems encountered during typical computer use. While this work is still preliminary, we believe it represents a step towards more self-reflective and self-repairing agent systems.

Abstract — This paper presents an imitation learning system capable of learning tasks in a complex dynamic real-time environment. In this paper we argue that social learning should be thought of as a special case of general skill learning, and that the biases it presents to the skill learning problem radically simplify learning for species with sufficient innate predisposition to harness this power. We decompose skill learning into four sub-problems, then show how a modification of Roy’s CELL system [1] can address all these problems simultaneously. Our system is demonstrated working in the domain of a real-time VR game, Unreal Tournament. Index Terms — Imitation, Social Learning, Memetics, Language.

We have developed a process model of problem solving with a simple control panel device. The model accounts well for many aggregate measures, including those from a study reported here (N=10): problem solving strategy, average fault-finding time, and the relative difficulty of faults. To further test the model, we compared the model's sequential predictions -- the order and relative speed that it examined interface objects and answered with a subject solving five tasks, making it one of the first models to have its sequential predictions compared with human data as they both learn. We found that the predicted actions matched and mismatch in systematic ways. The correspondences showed that: (a) subjects were reflecting or checking their work; suggesting that learning mechanisms can, in some instances, model the speed of learning while not completely modelling the mechanism; (b) mouse movements in the interface provide support for the model's predictions of what the subject attended to; providing further evidence that protocols can be augmented with mouse traces; and (c) while the comparison of sequential predictions may be becoming more tractable, the effects of learning raises new difficulties, such as assigning credit to model structures that change with performance.

Artificial cognitive systems sometimes neglect the impact on action selection of natural durativestate mechanisms like emotions and drives. These chemically-regulated motivation systems assist natural action selection through temporarily focusing an agent’s behavioural attention on particular problems. This can improve efficiency by avoiding dithering, but taken to extremes can be inefficient in ways that seem cognitively improbable for mammal-level intelligence. This article demonstrates a flexible latching method that provides appraisal-based sensitivity to interruption, allowing reassessment of the current focus of attention. This drastically improves efficiency in handling multiple competing goals at the cost of a surprisingly small amount of extra cognitive complexity.

Chemical / hormonal diffusion is the phylogenetically oldest form of biological action selection. In this paper we argue its persistence in higher animals is a consequence of its utility in solving problems of dithering between high-level goals. Chemical state underlying emotions and drives provides greater persistence more easily than the electrical action potential systems underlying the fine details of action sequencing, while also providing periodicity and transience not easily afforded in longer-term learning systems such as synaptic plasticity. We argue that artificial real-time autonomous systems require similar systems, and review our own efforts and approaches to providing these.

It has been suggested that the enterprise of developing mechanistic theories of the human cognitive architecture is flawed because the theories produced are not directly falsifiable. Newell attempted to sidestep this criticism by arguing for a Lakatosian model of scientific progress in which cognitive architectures should be understood as theories that develop over time. However, Newell’s own candidate cognitive architecture adhered only loosely to Lakatosian principles. This paper reconsiders the role of falsification and the potential utility of Lakatosian principles in the development of cognitive architectures. It is argued that a lack of direct falsifiability need not undermine the scientific development of a cognitive architecture if broadly Lakatosian principles are adopted. Moreover, it is demonstrated that the Lakatosian concepts of positive and negative heuristics for theory development and of general heuristic power offer methods for guiding the development of an architecture and for evaluating the contribution and potential of an architecture’s research program.

Multimodal interfaces are becoming increasingly important with the advent of mobile devices, accessibility considerations, and novel software technologies that combine diverse interaction media. This thesis investigates systems support for web browsing in a multimodal interface. Specifically, we outline the design and implementation of a software framework that integrates hyperlink and voice interaction. This enables the user to engage in out-of-turn interactions to personalize access at an information site. For the developer, the framework enables the creation of sites that adapt to the needs of users, yet permits fine-grained control over what interactions to support. Design methodology, implementation details, and two case studies are presented.

As cognitive architectures become ever more ambitious in the range of phenomena they are to assist in producing and modeling, there is increasing pressure for diversity in the mechanisms they embody. Yet uniformity remains critical for both elegance and extensibility. Here, the search for uniformity is continued, but shifted downwards in the cognitive hierarchy to the implementation level. Factor graphs are explored as a promising core, with initial steps towards a reimplementation of Soar. The ultimate aim is a uniform implementation level for cognitive architectures affording both heightened elegance and expanded coverage.