The immediate goal of this work was to develop a tool that:

1. Performs rapid recognitional inferences and planning within a large (expert) belief network
2. Exemplifies human limitations on computational resources and attention
3. Uses reflective strategies to help overcome computational limitations and deal with uncertainty.

Such a tool would be capable of simulating highly proficient, subtle, and creative aspects of human decision making in real-world domains.

Proficient decision makers, in relatively familiar situations, rapidly settle on situation interpretations and plans in the face of new observations and changing goals. This kind of situaion understanding and planning can be largely understood in terms of causal knowledge structures that we call mental models.

The reflexive subsystem shows how mental models can be generated and maintained in a dynamic situation. The starting point of the reflexive subsystem was a system called Shruti, developed by Professor Lokendra Shastri, of the International Computer Science Institute, University of California at Berkeley (Shastri & Ajjanagadde, 1993). Dr. Shastri is a long-time collaborator of CTI.

Shruti is unique in combining speed and scalability, with representation of crucial relational aspects of real-world decision making. To accomplish speed and scalability, Shruti utilizes parallel, connectionist processing. To keep track of relational reasoning, Shruti uses temporal synchrony of firing in nodes throughout the network to represent information about the same object. (Most connectionist models represent fuzzy similarity relations that blur the way objects interact. Symbolic architectures easily represent relational facts, but are usually slow and non-scalable.)

In Shruti, activation goes out from nodes that represent either sensory or linguistic inputs and/or an internally generated question to be settled, and returns when circuit is completed by other nodes that are linked to them. Shruti represents a situation model and/or decision by the emergence of a stable activation cycle within the network.

Extensions to Shruti were necessary both to improve its representation of reflexive reasoning and to make it work in conjunction with a reflective subsystem. Among the extensions that we have developed are:

1. Propagation of utility as well as belief, so that Shruti settles on actions at the same time as it settles on a situation interpretation.
2. Reasoning both backward, to find explanations, and forward, to generate predictions.
3. Mechanisms required for shifting attention, such as (a) priming effects that temporarily store results through a series of attentional shifts, and (b) long-term storage of aggregated results at the edge of the currently active part of the network.
4. Tuning of link strengths through backpropagation.

Attentional limits on dynamic access to long term memory (LTM) emerge naturally from the computational structure of Shruti and the neuro-biological constraints it respects. These inmply that not all information known by the agent can be brought to bear on a problem at the same time by purely reflexive processes.

The reflective subsystem critiques the conclusions of reflexive processing and guides its subsequent progress. (See the Recognition / Metacognition model) Features of the reflective subsystem include:

1. Methods for identifying qualitatively different types of uncertainty based on activation patterns in the reflexive system
2. Methods for identifying beliefs most likely to be responsible for different types of uncertainty
3. Strategies for shifting attention to beliefs most likely to be responsible for uncertainty. The results of these attentional shifts is the activation of previously dormant information in long-term memory that is likely to help resolve the uncertainty.

The metacognitive system learns to combine a set of simple operations: inhibiting recognitional responding, activation of new information internally by shifting focal attention, and clamping truth values (which is itself a form of persistent attention to a node). These operations are simple and both psychologically and biologically plausible.

The metacognitive system learns to combine these operations in response to different patterns of uncertainty by reinforcement and associative learning processes. Through such learning, the metacognitive system acquires a rich repertoire of uncertainty handling strategies. These strategies include both domain-specific and general elements. The result is a dynamic process of evaluating and improving mental models of the situation and plans.