Learning via Gradient Descent in Sigma

Integrating a gradient-descent learning mechanism at the core of the graphical models upon which the Sigma cognitive architecture/system is built yields learning behaviors that span important forms of both procedural learning (e.g., action and reinforcement learning) and declarative learning (e.g., supervised and unsupervised concept formation), plus several additional forms of learning (e.g., distribution tracking and map learning) relevant to cognitive systems/modeling. The core result presented here is this breadth of cognitive learning behaviors that is producible in this uniform manner. One of the key hypotheses investigated during Soar's early years was that a simple learning mechanism - chunking - when integrated into an appropriate architecture could yield a general learning mechanism capable of acquiring the diversity of knowledge required by a cognitive system (Laird, Rosenbloom & Newell 1986). Although this proved to be a bridge too far - with Soar later to incorporate additional reinforcement, episodic and semantic learning mechanisms (Laird, 2012) - much was learned in exploring this hypothesis over the years (Rosenbloom, 2006). Despite the limitations eventually evident in Soar, the drive towards general learning mechanisms in cognitive architectures/systems remains appealing. To the extent it is feasible, it yields deeper and more elegant theories of intelligent behavior with broader scientific reach (Deutsch, 2012). Here we report results from a somewhat more modest such effort that, like Soar, is based on integrating a simple learning mechanism into an appropriate architecture; however, the particular mechanism here - a local, online variant of gradient descent - is quite different from chunking, and is integrated into a hybrid mixed architecture called Sigma (Σ). The results are interestingly different from those obtained with chunking in Soar.