The Sigma Cognitive Architecture and System

Sigma (Σ) is a nascent cognitive system – an integrated computational model of intelligent behavior, whether natural and/or artificial – that is based on a novel cognitive architecture: a model of the fixed structure underlying a cognitive system [1]. The core idea behind Sigma is to leverage graphical models [2, 3] – with their ability to yield state-of-the-art algorithms across the processing of signals, probabilities and symbols from a single representation and inference algorithm – in constructing a cognitive architecture/system that meets three general desiderata: grand unification, functional elegance and sufficient efficiency. A unified cognitive architecture traditionally attempts to integrate together the complementary cognitive capabilities required for human(level) intelligent behavior, with appropriate sharing of knowledge, skills and uncertainty among them. A grand unified architecture goes beyond this, in analogy to a grand unified theory in physics, to include the crucial pieces missing from a purely cognitive theory, such as perception, motor control, and emotion. This shifts issues of embodiment, grounding and interaction into the foreground, to converge with work on robot and virtual human architectures, but without then relegating traditional cognitive concerns to the background. Sigma approaches grand unification via a hybrid (discrete + continuous) mixed (symbolic + probabilistic) combination of: (1) graphical models, in particular factor graphs with the summary product algorithm [4]; plus (2) piecewise-continuous functions [5]. Functional elegance implies combining broad functionality – grand unification in this case – with simplicity and theoretical elegance. The goal is something like a set of cognitive Newton's laws that yield the required diversity of behavior from interactions among a small set of general primitives. If the primitives are combinable in a flexible enough manner, new capabilities continue to flower without the need for new modules; and integration occurs naturally through shared primitives. The Soar architecture, many of whose strengths and weaknesses have inspired choices in Sigma, took a similar approach in its early years [6], while AIXI can be seen as striving for an extreme version of it [7]. Although doubts remain as to whether natural cognitive systems are elegant in this manner – as opposed to mere evolutionary patchworks – and whether such elegance is even computationally feasible in artificial cognitive systems, developments such as rational analysis on the natural side [8] and graphical models on the artificial side provide continued promise. Despite the questions, functional elegance maintains its allure because, if it is in fact achievable, it should …