Emotions, Motivation-Based Action Selection and Dynamic Environments

Abstract

In contrast to traditional approaches, where the focus is on developing or evolving artificial “brains” as the route to artificial intelligence (AI) more recent approaches have increasingly emphasised and modelled the role of “bodies” and “environments”. In turn, this has further encouraged ideas regarding aspects of intelligence as being best thought of as distributed across agent brains, bodies and environments. That is, as system properties emerging from interactions of these components. Action selection is commonly recognised as one of the problems all agents, whether biological or artificial, must face: deciding at any given moment “what to do next”. Researchers have generated many different action selection mechanisms as “solutions” to this problem. However, in the work of this thesis, we focus on one which takes its inspiration from biological ideas about the role and possible neural substrates of emotion. We use this to consider how models of brain-body-environment interactions might be more useful for the study of emotion, as well as action selection mechanisms. For, despite the many mechanisms proposed, the literature still lacks systematic ways to analyse their performance in combination with different physical and/or perceptual capabilities. That is, factors relating more directly to agent embodiment. In this thesis we have studied the performance of our selected architecture in a robotic predator-prey scenario known as the Hazardous Three Resource Problem. The predator-prey relationship is popular in artificial intelligence, both as an action selection problem and a situation which enables study of agent-agent interactions. Predators can act as catalysts for the evolution of prey agents in a “survival of the fittest” sense while, in their turn, prey agents are tests of predator ingenuity. For us, however, it is also a situation where emotion might naturally be assumed to have useful functions. To study action selection, emotion and brain-body-environment interactions in an artificial predator-prey relationship, we both advocate and adopt a bottom-up, animat approach. The animat approach to AI is one that emphasizes characteristics neglected by more traditional approaches. As such, it has embraced the study of robotic agents. One reason for this is the process of designing “real-world” agents forces us to consider practicalities simulations might not. What makes the use of robots particularly appealing for our work, however, is how it can give us a greater appreciation of more physical aspects of intelligence such as agent morphology and its integration with agent control mechanisms as well as environmental dynamics. Using LEGO robots, we show how the performance of our architecture varies in our chosen scenario with aspects of agent brain, body and environment. We argue our results complement existing research by contributing evidence from a real-world implementation, explicitly modelling ideas about action selection and emotion as distributed across, or best thought of as emerging from interactions between, agent brain, body and environment. In particular, this thesis shows how our selected architecture varies and benefits from further integration with aspects of agent “body”. It also acts as an example of an alternative form for the bottom-up development of artificial emotion, demonstrating wider applications for creating more adaptive action selection mechanisms. Comparing the robotic predator-prey relationships we have created to ethological evidence and theories, we argue our architecture may also have specific potential for future research and applications — having already proven itself capable of emerging multiple functions and properties