Argo: an analogical reasoning system for solving design problems

The static and predetermined capabilities of many knowledge based design sys tems prevent them from acquiring design experience for future use To overcome this limitation techniques for reasoning and learning by analogy that can aid the design process have been developed These techniques along with a nonmonotonic reason ing capability have been incorporated into Argo a tool for building knowledge based systems Closely integrated into Argo s analogical reasoning facilities are modules for the acquisition storage retrieval evaluation and application of previous expe rience Problem solving experience is acquired in the form of problem solving plans represented as rule dependency graphs From increasingly abstract versions of these graphs Argo calculates sets of macrorules These macrorules are partially ordered according to an abstraction relation for plans from which the system can e ciently retrieve the most speci c plan applicable for solving a new problem Knowledge based applications written in Argo can use these plan abstractions to solve problems that are not necessarily identical but just analogous to those solved previously Ex periments with an application for designing VLSI digital circuits are yielding insights into how design tools can improve their capabilities as they are used Introduction and Background A number of knowledge based systems for design have been developed recently These systems are particularly suited to situations in which heuristic expert knowledge must be employed because algorithmic techniques are unavailable or prohibitively expensive Restrictions on the types of problems or domains handled by such design systems are progressively being eased Unfortunately the knowledge embodied in many of these systems is static it fails to capture the iterative aspects of the design process that involve solving new problems by building upon the experience of previous design e orts Given the same problem ten times these systems will solve it the same way each time taking as long for the tenth as for the rst The work reported here is based on the contention that a truly intelligent design system should improve as it is used i e it should have the means for remembering the relevant parts of previous design e orts and be able to employ this accumulated experience in solving future design problems Learning from experience is a powerful technique used by humans to improve their problem solving ability For a design tool the remembered experience should consist of design results design plans and preferences among these results and plans These constitute di erent aspects of previous design e orts that the design tool can use as training examples Learning from Experience Existing approaches to learning from experience attempt to generalize these training exam ples in order to obtain more widely applicable results The STRIPS problem solving system incorporates a technique for generalizing plans and their preconditions based on the formation of macro operators MACROPs In this technique an existing plan consisting of a sequence of operators whose execution yields a goal state is stored in a data structure called a triangle table This table represents the preconditions and postconditions for each operator in the plan The plan is generalized by replacing all precondition constants by distinct parameters and then correcting for overgeneralization by substituting for incon sistent parameters The resultant generalized plan a MACROP is stored and later used as either a plan a set of subplans or an execution monitor A better procedure for generalization developed in the context of learning from exam ples uses a proof based explanation or veri cation mechanism often termed explanation based generalization EBG It is an improvement over the use of a triangle table in that it does not require any heuristics to compensate for possible over generalizations The proof employed comprises information about why a training example satis es a particular goal The procedure involves rst a modi ed regression of the goal through the proof structure whereby su cient constraints on the domain of train ing examples for which the proof holds are computed These constraints are based on the codomain of goals allowed The second stage of the procedure is to reapply the proof structure to the resultant generalized domain to obtain a generalized codomain In the terminology used above a plan is like a proof a plan precondition is the domain for the proof and the resultant design is the codomain For design problems EBG like generalizations are limited in that they arbitrarily give equal weight to all portions of the examples without regard to whether each portion is relevant or important to solving future problems More abstract generalizations can be obtained by taking this factor into account Abstract planning i e choosing a partial sequence of operators to reach a goal is accomplished in ABSTRIPS by ignoring operator preconditions considered to be details Criticality values are attached to the preconditions of operators to determine their importance These values are computed based on the presumed di culty of satisfying each precondition if it is not already satis ed Only if a plan succeeds at an abstract level is it expanded by the addition of subplans to handle the details at a subsequent level Another technique for reusing past design experience is to replay a previously recorded plan or design history This approach is interesting in its exibility with respect to replaying portions of a stored plan to solve or at least partially solve a new problem Unfortunately the correspondence between the stored plan and subproblems of a partial design is di cult to establish automatically The transfer of experience from previous problem solving e orts to new problems has also been accomplished via analogical reasoning methods Analogical reason ing is a mapping from a base domain to a target domain that allows the sharing of features between these domains With respect to problem solving many of the previously reported methods are limited by their requirements that either new problems be very similar to previously solved ones or analogies be supplied by a user and match perfectly