The work of Soloway on a system designed to learn some of the concepts of the game of baseball will provide an example of what is often termed knowledge based learning. The dominant approach to induction or learning was to attempt to develop learners that began with the minimum knowledge possible. The work by Mitchell on learning in version spaces can be viewed as a definition of a minimally "knowledge-biased" learner that still guaranteed both coverage of the training sequence and prediction about future training examples.

     A bias remained because some descriptive language had to be chosen within which to describe the examples and the concepts which covered the training examples. Thus, it was clear that what we have termed the representational bias remains unless we can:

1. agree on a single representation language; and
2. all concepts can be defined in terms of this single representation language.

Knowledge and Multiple Representations

     Games are artifacts. That is, someone invented the games that we play. And, it is quite apparent that the terms that we use to describe a particular game are related to each other and are unique to that particular game. For example, if you and I are playing tennis, I may swing and miss a particularly well-placed slice serve wide to my forehand. Or we may be playing baseball and I swing and miss a particularly well-thrown curve that breaks away from me on the outside corner of the plate. In both cases, I swung and missed a ball that was moving with considerable speed and breaking away from my body. In the first case, a "point" and in the second case a "strike" is tallied against me. The similarity of physical description does not signal that these two instances are examples of the same concept. In fact they are quite different things despite the physical similarity.

     Notice that we have implicitly introduced the idea that a particular set of events or observations can be described in a physical language as well as the language of the game. And, in general we distinguish between the physical activities that someone does and the plans that those activities might be part of. This distinction is most strikingly obvious when we observe that doing nothing at the physical level can be described as doing something at the plan level, and, conversely, doing something at the physical level may correspond to doing nothing at the plan level. For example, in baseball if the ball is pitched and even though I do not swing at the pitch it may still be appropriately described as a 'strike.' Or, swinging prior to the pitcher throwing the pitch is the same as not swinging prior to the pitcher throwing the pitch -- in both cases I am 'at bat waiting for the next pitch.'

     These examples suggest that the unsupervised learning of these types of concepts might be extremely difficult and perhaps impossible without an appropriate knowledge-based bias. We can develop a better understanding of these issues by examining the Baseball learner that was designed by Eliot Soloway. This research is described in Soloway, E., BASEBALL: An example of knowledge-directed machine learning. In Bower, G. H. (Ed.), The psychology of learning and motivation, Vol. 20. New York: Academic Press Inc., 1986. pp. 193-236. The presentation of this work here is based both on this article and on Soloway's dissertation.

     This is one of the possible ways in which to design a knowledge-based learner of this type. Further research may inform us concerning the relative advantages or disadvantages of this particular architecture. Examining this architecture in some detail will help to clarify some of the problems that any learning architecture of this type must address.

     Probably the most salient characteristic of this type of learning is the necessity to form the hypotheses in a language that is different from the observations. We will refer to these two languages as the Observation Language and the Plan Language. It is assumed that the learner is initially totally ignorant about the game of baseball. However, the learner is assumed to be totally competent at recognizing and representing events that occur in the physical world. 'Balls moving through the air', 'persons swinging bats,' ...'throwing balls' and so on are assumed to be already learned concepts. The figure below provides a more formal characterization of the learner's knowledge of the physical world of actions.

  The types of expressions that the learner initially knows and represents are shown in parentheses. There are eleven types and the capitalized terms (e.g., HOLDOBJECT, RUN, AT,...) represent the relation and the lower-cased terms the types of parameters or terms (e.g., time, location, ...) that are associatred with the relation. The values of the types of parameters are shown in the gray box. Note that time is totally ordered using the integers as the index.

     A program was written to generate the input....the activities of a baseball game described in these expressions. The figure below shows a set of expressions for three time periods, namely; 102, 103, and 104. A collection of expressions that are all indexed at the same time index is referred to as a snapshot.

Three main processing levels can be distinguished. These are:

• ATTENTION FOCUSING (AF) which uses knowledge that change and high energy activity are important in physical games. This assumption is used to filter the snapshots and this filtering serves to focus/bias the subsequent learning. AF maps from a linear sequence of completely described snapshots to a structured order of partially described snapshots where the structure is a selected partition called episodes which yields an activity cycle structure over the input sequence. The physical enablement relations between actions is used to compute causal enablement chains which provide a basis for the identification of an activity cycle. Note that episodes are described in the Observation Language.

Example of a Physical Enablement Inference

• INTERPRETATION which uses knowledge about cooperative and competitive plans as a basis for taking the output of the lower level and reformulates this output in the language of plans. Goals of the players, the cooperative or competitive structure of the interaction, and the success or failure of the goals. Knowledge of four specific types of competitive interactions can be distinguished:
  • Physical Competition, (e.g. pitcher/batter);
  • Order of occurrence, (e.g. beating out an infield ground ball);
  • State-of-Distinguished-Object, (e.g., hit ball caught in air);
  • Logical-Competition ,(e.g., called strike, this is used to force correctness with respect to the assumption that members of the same team are in a cooperative plan and in a competitive plan with member of the other team.)
    
• EVALUATION which determines whether the set of current hypothesis are internally consistent and overall consistent with the assumptions involved in competition and cooperation. Generalization of hypothesis was designed to be conservative because one of the problems of this type of learning is overgeneralization - particularly because there is no teacher.