All intelligence relies on search --- for example, the search for an intelligent agent's next action. Search is only likely to succeed in resource-bounded agents if they have already been biased towards finding the right answer. In artificial agents, the primary source of bias is engineering. This dissertation

Hierarchical organization has become an unfashionable model of intelligent control within some communities of both natural and artificial intelligence. What has replaced it are models based on parallel distributed processes, both neural and behavior based, or dynamical systems theory, which denies modularity, let alone rigorous structure. In this paper we present experimental results demonstrating an artificial reactive hierarchybased system that outperforms fully parallel systems in a highly dynamic environment with a large number of conflicting goals. This work is conducted in Tyrrell’s (1993) Simulated Environment and can be seen as an extension of his work on comparing action selection mechanisms. We observe that the hierarchical strategy has also been well demonstrated in nature. We argue that, for complex intelligences, preserving full reactivity may not be worth the cost in terms of the complexity of action selection.

Abstract. This volume is intended to help advance the field of artificial neural networks along the lines of complexity present in animal brains. In particular, we are interested in examining the biological phenomena of modularity and specialized learning. These topics are already the subject of research in another area of artificial intelligence. The design of complete autonomous agents (CAA), such as mobile robots or virtual reality characters, has been dominated by modular architectures and context-driven action selection and learning. In this chapter, we help bridge the gap from neuroscience to artificial neural networks (ANN) by incorporating CAA. We do this both directly, by using CAA as a metaphor to consider requirements for ANN, and indirectly, by using CAA research to better understand and model neuroscience. We discuss the strengths and the limitations of these forms of modeling, and propose as future work extensions to CAA inspired by neuroscience.

One of the greatest obstacles to designing a mind is the complexity of integrating different process types, time frames and representational structures. This paper describes a methodology for addressing this obstacle, Behavior Oriented Designed (BOD), and explains it in the context of creating an agent capable of natural language dialogue. 1

: `Pigs and people' is a simulated environment, in which action selection mechanisms can be evaluated and compared. Action selection mechanisms attempt to solve the action selection problem faced by both animals and robots: the problem of selecting which actions to perform in order to achieve heterogeneous and possibly conflicting goals. In this project, two action selection mechanisms are implemented: the non-learning drives mechanism, and W-learning. The non-learning mechanism considerably outperforms the learning mechanism. The results achieved by the two mechanisms are compared and analysed in an attempt to explain the difference in performance.

Forty-two! yelled Loonquawl. Is that all you ve got to show for seven and a half million years w ork? I checked it very thoroughly, said the computer, and that quite definitely is the answer. I think the problem, to be quite honest with you, is that you ve never actually known what the question is. But it was the Great Question! The Ultimate Question of Life, the Universe and Everything, howled Loonquawl. Yes, sai d Deep Thou ght wit h th e air of on e wh o suffers fools gla dly, but what a ctually is it? Douglas Adams, The Hitchhikers Guide to the Galaxy 1

Abstract. We present a schema-based agent architecture which is inspired by an ethological model of the praying mantis. It includes an inner state, perceptual and motor schemas, several routines, a fovea and a motor. We describe the design and implementation of the architecture and we use it for comparing two models: the former uses reactive, stimulusresponse schemas; the latter involves also forward models, which are used by the schemas for generating predictions. Our results show an advantage in using anticipatory components inside the schemas 1. 1

This is an extended abstract describing work in progress. We are developing an artificial life (ALife) model the various sorts of social behaviors displayed by colonies of non-human primates. We hope to use this ALife model to support work on an ethological model also under development which examines the relationship between conflict management behaviors displayed by a species and the dominance relationships between individuals of that species. In this paper, we describe the relationship between emotions and action selection in our ALife model.

This dissertation introduces PyPOSH, a reactive agent architecture with loadable behavioral modules based on Bryson’s Parallel-Rooted Ordered Slip-Stack Hierarchical action selection model. The framework utilizes a modular and object-oriented interface to behaviors built utilizing agile and component-based methods.

There is strong evidence in the literature for at least three di#erent representations underlying transitive choice in various species of animals. This paper focuses primarily on understanding one of the most neglected: the production-rule model of Harris and McGonigle (1994). The production-rule model has been to date the best model at accounting for the performance of squirrel monkeys (Saimiri sciureus) and human children under 7 trained on the transitive inference (TI) task (e.g. given A > B and B > C, then A > C) when presented with three items from the training set. This paper presents a new neurologically-plausible version of this representation, the two-tier model, which explains how this sort transitive performance is learned. This new model perfectly replicates the positive aspects of the productionrule model while accounting for more of the data, particularly subject's failures to learn transitive inference. This paper also discusses how the two-tier model fits with other transitive inference models, and characterises how to recognise which TI representation underlies which sorts of TI performance. Of the three representations discussed, we suggest that the two-tier model may be the most relevant for understanding general-purpose primate task learning, and that it may even provide the scaffolding for the human acquisition of concrete operational thought.