TETON (Vanlehn, K. , Ball, W.)


Philosophy

A unified theory of cognition should be developed piecewise, by modeling specific empirically verifyiable behaviors and specifying how they relate to each other to form a cognitive architecture. This methodology leads to a development of models of higher level cognitive functions first, such as goal reconstruction, which are then mapped explicitly onto lower level functions such as memory, a attention, and motor control.

Architecture:

TETON's architecture is composed of three parts, a long term memory, a working or short term memory , and an execution cycle. The long term memory stores procedures (operators) and selection rules. The working memory is itself comprised of four parts, but its main use is as storage for recent results of the execution cycle, which are goals. The execution cycle uses selection rules in the long term memory to choose an operator. This operator is then interpreted and it's result placed in working memory. The architecture will automatically subgoal if an impass is reached (no operator is preferred , and there are multiple methods for resolving such impasses.

The long term memory is a knowledge base which is used to store operators, selection rules, and general facts. An operator has some selection critera, and a set of preconditions. If the preconditions are met, then the body is interpreted, and can effect working memory elemetns or may delibertly create subgoals. Operators also contain 'shortcuts' which inform the interpreter that his goal has already been achieved. This situation occurrs during 'goal - reconstruction', which happens when one or more goals are forgotton. When it does occurr, it allows TETON to simply mark the goal as completed without executing it. The other type of knowledge in the long term memory are selection rules. Selection rules are of three forms, Consideration, Rejection, and Preference. They operate, as thier names suggest, by forming a set of rules to consider, rejecting some, and creating a partial order of the rest. This is very similar to the way Soar (J.Laird)) choses operations.

The main working memory contains goals produced by recent computations Goals in TETON are slot based structures which can represent past, present, and future goals. They also contain the state to be achieved, and a link to the state they were created from. Other information may also be present - dependant on the needs of the procedures in long term memory. There are three other divisons of working memory besided main, including a situation memory which is a representation of the external environment, a limited capacity buffer which stores recent important events, and a scratchpad which is used for lookahead search in planning actions. Theor

The execution cycle first decides what to do by interpreting selection rules, then preforms the operation by interpreting body of the operator selected. The execution cycle will automaticlly create an impass if no operator is uniquely selected. Impasses are resolved by multiple methods, including psychologicaly plausable ones such as simply skipping it or randomly choosing an operator. One important property of this selection method is that it is NON-LIFO, in other words any current goal can be selected for achievement , not simply the last goal created. The execution cycle also checks to see if the current goal has a shortcut condition. The presence of a shortcut condition is a flag that the goal has already been achieved, and that what the executor is now working on is what is called 'goal-reconstruction', because one or more previously accomplished goals have been forgotton. In this case the goals do not need to be executed and can simply be marked as completed.

Agent Properties:

TETON's knowledge is stored mainly as procedural knowledge in it's long term memory component. The long term memory component is impenertrable to the rest of the system, and the representation language used is not very usefull for representing declaritive knowledge, but nothing prohibits such knowledge from being incorporated in the future. The long term memory component also contains strategic knowlege which is used to impose a partial ordering on the set of procedures selected during the execution cycle in an attempt to determine the sole best procedure to execute in the current situation.

Capabilities:

TETON is basically a problem solving and planning architecture. Learning is not incorporated currently, but nothing inherent to the architecture prohibits it's inclusion. One of the authors main points was the incorporation of 'goal reconstruction' into TETONS problem solving abilities. Goal reconstruction is the process of rebuilding lost goal structures through the incorporation of some external notation which acts as a reminder of previously achieved goals. The process is included for phychological validity, in terms of limited short term memory and opportunistic problem solving. TETON's short term memory is limited, and therefore 'forgets' some subgoals if distracted for some time. The system then has the ability to reconstruct these forgotton goals given a top level goal and utilize the shortcut operator slot to avoid re - executing these goals. The same reconstruction method applies to problems which are given already partially solved, the system will attempt to resolve the entire problem, but those goals which are already completed in the environmnent are 'shortcut'.

Environment and Agent Body:

The problems TETON has been exposed to are simple, static subtraction problems. The problems are presented visually, and the goal of the system is to solve the problem using methods similar to those used by humans. This includes utilizing the environment as a means of representing aspects of the problem it does not have the ability to represent internally. Currently, in TETON, the external representation is limited to information regarding the partial completeness of a problem. TETON assumes a perfect sensing of the environment as presented, and is not concerned with speed of response. The architecture however seems to impose no restrictions on Scalability in either of these areas.

Issues:

One of the main limitation of the TETON architecture currently is its focus on Procedural knowledge. This effects its scalability because for Procedures to be effective in diverse environments, the must have access to larger and larger amounts of declaritive knowlege, or facts. TETON however is meant as more of a Psychologically plausable computational model of human cognition, in particular, the phycological ramifications of goal reconstructuction are explored and empirically tested.

Theory:

The environment, along with our limited capacity to represent information determines the algorithm we use to solve mathematical problems.
Other Architectures

Back to the Title Page