Research on cognitive architectures attempts to develop unified theories of the mind. This paradigm incorporates many ideas from other parts of AI, but it differs enough in its aims and methods that it merits separate treatment. In this paper, we review the notion of cognitive architectures and some recurring themes in their study. Next we examine the substantial progress made by the subfield over the past 40 years, after which we turn to some topics that have received little attention and that pose challenges for the research community.
We describe a generic approach for modeling the impact of emotion on cognition, perception, and behavior. The approach can model the effects of transient emotional states, longer moods, and stable personality and temperamental factors. The underlying assumption is that one of the primary ways in which emotions influence cognition and perception is by modulating a variety of processing parameters. We illustrate the approach in the context of both a generic integrated architecture of cognition, and a specific architecture, currently under development, designed to model decision making behavior. In this context, we illustrate how the approach would be instantiated within several representational formalisms (e.g., rules, belief nets). We focus on modeling the impact on tactical decision-making of three specific emotional states that have been studied extensively in experimental psychology: anxiety, negative affect (e.g., depression), and obsessiveness. The proposed approach can then be used both for investigating the interaction between cognition and emotion, and the resulting behavior, and for modeling specific types of personalities in interactive environments.
Historically, AI research has understandably focused on those aspects of cognition that distinguish humans from other animals - in particular, our capacity for complex problem solving. However, with a few notable exceptions, narratives in popular media generally focus on those aspects of human experience that we share with other social animals: attachment, mating and child rearing, violence, group affiliation, and inter-group and inter-individual conflict. Moreover, the stories we tell often focus on the ways in which these processes break down. In this paper, I will argue that current agent architectures don't offer particularly good models of these phenomena, and discuss specific phenomena that I think it would be illuminating to understand at a computational level.
A key challenge in creating simulated agents is to produce sufficiently realistic behavior. A critical component of such realism is the range of variations in behaviors exhibited by humans. Whether these be'leaps of genius', surprising reactions, specific biases, suboptimal behaviors, or simply errors, these inconsistencies and idiosyncracies are quintessential human qualities. These variations are due to a variety of factors, including varying levels of intelligence and skill, differences in cognitive and decision making styles, personality differences, and differences in specific affective states and moods. Collectively, these factors are termed individual differences.
The anthropological and economic history of humanity gives evidence of a progression of cognitive frameworks. There are three cognitive perspectives, in order: living in the present, living in the past, and living in the future. They correspond to three levels of competency with abstract thought: concrete thought only, abstract thought with correlations, and abstract thought with both correlations and causality. This appears to explain the fundamental differences between primitive cultures, traditional cultures, and modern cultures: differences in economics, politics, personality, and anthropological differences in general. So, not only does this theory succinctly explain a wide range of human behavior, but because it does, it appears to be a valid theory and a promising way to decompose abstract thought into its component parts for future cognitive research. These frameworks are discussed along with their implications of exploiting this progression to simplify the problem of developing an AI.