Susan L. Epstein

For the right reasons: The FORR architecture for learning in a skill domain

The theme of this article is that knowledge acquisition can be managed as a transition from general expertise to specific expertise. FORR is a general architecture for learning and problem solving that models expertise at a set of related problem classes. This architecture postulates initial brood domain knowledge, and gradually specializes it to simulate expertise in individual problem classes. FORR is based upon a realistic portrayal of the nature of human expertise and its application. Rather than restrict learning to a single method or a single kind of knowledge, the architecture pragmatically requires multiple, disagreeing heuristic agents to collaborate on decisions. A FORR-based program learns both from its apprenticeship to an external expert model and from practice in its domain. An implementation for game playing is described that raises interesting issues about the organization and modification of conflicting expertise, and the role that experience plays in such learning. FORRs principal strengths are its smooth integration of multiple expertise, its ability to learn many ways, its tolerance for human and machine error, its graceful degradation, its transparency, and its support for a developmental paradigm.

Toward an Ideal Trainer

This paper demonstrates how the nature of the opposition during training affects learning to play two-person, perfect information board games. It considers different kinds of competitive training, the impact of trainer error, appropriate metrics for post-training performance measurement, and the ways those metrics can be applied. The results suggest that teaching a program by leading it repeatedly through the same restricted paths, albeit high quality ones, is overly narrow preparation for the variations that appear in real-world experience. The results also demonstrate that variety introduced into training by random choice is unreliable preparation, and that a program that directs its own training may overlook important situations. The results argue for a broad variety of training experience with play at many levels. This variety may either be inherent in the game or introduced deliberately into the training. Lesson and practice training, a blend of expert guidance and knowledge-based, self-directed elaboration, is shown to be particularly effective for learning during competition.This paper demonstrates how the nature of the opposition during training affects learning to play two-person, perfect information board games. It considers different kinds of competitive training, the impact of trainer error, appropriate metrics for post-training performance measurement, and the ways those metrics can be applied. The results suggest that teaching a program by leading it repeatedly through the same restricted paths, albeit high quality ones, is overly narrow preparation for the variations that appear in real-world experience. The results also demonstrate that variety introduced into training by random choice is unreliable preparation, and that a program that directs its own training may overlook important situations. The results argue for a broad variety of training experience with play at many levels. This variety may either be inherent in the game or introduced deliberately into the training. Lesson and practice training, a blend of expert guidance and knowledge-based, self-directed elaboration, is shown to be particularly effective for learning during competition.

Conserved Processes and Lineage-Specific Proteins in Fungal Cell Wall Evolution†

ABSTRACT The cell wall is a defining organelle that differentiates fungi from its sister clades in the opisthokont superkingdom. With a sensitive technique to align low-complexity protein sequences, we have identified 187 cell wall-related proteins in Saccharomyces cerevisiae and determined the presence or absence of homologs in 17 other fungal genomes. There were both conserved and lineage-specific cell wall proteins, and the degree of conservation was strongly correlated with protein function. Some functional classes were poorly conserved and lineage specific: adhesins, structural wall glycoprotein components, and unannotated open reading frames. These proteins are primarily those that are constituents of the walls themselves. On the other hand, glycosyl hydrolases and transferases, proteases, lipases, proteins in the glycosyl phosphatidyl-inositol-protein synthesis pathway, and chaperones were strongly conserved. Many of these proteins are also conserved in other eukaryotes and are associated with wall synthesis in plants. This gene conservation, along with known similarities in wall architecture, implies that the basic architecture of fungal walls is ancestral to the divergence of the ascomycetes and basidiomycetes. The contrasting lineage specificity of wall resident proteins implies diversification. Therefore, fungal cell walls consist of rapidly diversifying proteins that are assembled by the products of an ancestral and conserved set of genes.

LEARNING TO SUPPORT CONSTRAINT PROGRAMMERS

This paper describes the Adaptive Constraint Engine (ACE), an ambitious ongoing research project to support constraint programmers, both human and machine. The program begins with substantial knowledge about constraint satisfaction. The program harnesses a cognitively‐oriented architecture (FORR) to manage search heuristics and to learn new ones. ACE can transfer what it learns on simple problems to solve more difficult ones, and can readily export its knowledge to ordinary constraint solvers. It currently serves both as a learner and as a test bed for the constraint community.

Pragmatic navigation: reactivity, heuristics, and search

Abstract FORR (FOr the Right Reasons) is an architecture for learning and problem solving that integrates a possibly incomplete and overlapping set of solution methods to address complex problems. Each method, although it represents some facet of domain expertise, may vary in reliability and speed. The principal contribution of this paper is the extension of FORR to include situation-based behavior (the serial testing of known, triggered techniques for problem solving in a domain) with reactivity and heuristic reasoning. FORR categorizes methods as reactive, heuristic, or situationbased, and addresses problem solving with one category of methods at a time. A hierarchical reasoner first has the opportunity to react correctly. If no ready reaction is computed, the reasoner activates a set of reactive triggers for time-limited search procedures tailored to specific situations. If they, too, fail to produce a response, the reasoner resorts to collaboration among heuristic rationales. All three components reference knowledge learned from experience. In a series of experiments, this architecture is shown to be effective and efficient. Ablation experiments demonstrate how each component plays an important role in problem solving. Additional contributions of this paper include a FORR-based, pragmatic, cognitively plausible approach to navigation with learned heuristic approximations that describe two-dimensional territory and travel experience through it, and a careful study of how situation-based behavior, reactivity, and heuristics interact there. Empirical evidence demonstrates that the resultant system is both effective and efficient, and guidelines for generalization to other domains are provided.

Prior knowledge strengthens learning to control search in weak theory domains

Rather than search exhaustively in large problem spaces, people use heuristics that entail minimal search and yet support intelligent decisions. Human performance in such spaces is robust and improves with experience. Game playing is a good example. the traditional AI approach to game playing, however, has been to build a program that plays only a single game, and plays it very well. Because exhaustive search of the game graph and minimax of the resultant values guarantees perfect play, programmed champions rely on deep, fast, efficient search. Despite their more limited, fallible memories and arguably slower processors, human masters often defeat a program whose search is incomplete. People search less, remember less, and somehow make the right choices. This article describes HOYLE, a program that learns to play specific games under the direction of a weak theory, clearly delineated prior knowledge about games in general. HOYLE differs from most game‐playing programs in two ways: it is able to play any of a broad class of games according to the rules, and it improves its performance through a variety of learning paradigms. HOYLE shows that, for a class of related problems, it is possible to replace deep search with a weak theory based upon heuristically selective cache memories and a consensus about action among reliable rationales.

Learning and discovery: one system's search for mathematical knowledge

The Graph Theorist, GT, is a system that performs mathematical research in graph theory. From the definitions in its input knowledge base, GT constructs examples of mathematical concepts, conjectures and proves mathematical theorems about concepts, and discovers new concepts. Discovery is driven both by examples and by definitional form. The discovery processes construct a semantic net that links all of GTs concepts together.

Learning Algorithm Portfolios for Parallel Execution

Portfolio-based solvers are both effective and robust, but their promise for parallel execution with constraint satisfaction solvers has received relatively little attention. This paper proposes an approach that constructs algorithm portfolios intended for parallel execution based on a combination of case-based reasoning, a greedy algorithm, and three heuristics. Empirical results show that this method is efficient, and can significantly improve performance with only a few additional processors. On problems from solver competitions, the resultant algorithm portfolios perform nearly as well as an oracle.

Composition-Modified Matrices Improve Identification of Homologs of Saccharomyces cerevisiae Low-Complexity Glycoproteins

ABSTRACT Yeast glycoproteins are representative of low-complexity sequences, those sequences rich in a few types of amino acids. Low-complexity protein sequences comprise more than 10% of the proteome but are poorly aligned by existing methods. Under default conditions, BLAST and FASTA use the scoring matrix BLOSUM62, which is optimized for sequences with diverse amino acid compositions. Because low-complexity sequences are rich in a few amino acids, these tools tend to align the most common residues in nonhomologous positions, thereby generating anomalously high scores, deviations from the expected extreme value distribution, and small e values. This anomalous scoring prevents BLOSUM62-based BLAST and FASTA from identifying correct homologs for proteins with low-complexity sequences, including Saccharomyces cerevisiae wall proteins. We have devised and empirically tested scoring matrices that compensate for the overrepresentation of some amino acids in any query sequence in different ways. These matrices were tested for sensitivity in finding true homologs, discrimination against nonhomologous and random sequences, conformance to the extreme value distribution, and accuracy of e values. Of the tested matrices, the two best matrices (called E and gtQ) gave reliable alignments in BLAST and FASTA searches, identified a consistent set of paralogs of the yeast cell wall test set proteins, and improved the consistency of secondary structure predictions for cell wall proteins.

PATTERN-BASED LEARNING AND SPATIALLY ORIENTED CONCEPT FORMATION IN A MULTI-AGENT, DECISION-MAKING EXPERT

As they gain expertise in problem solving, people increasingly rely on patterns and spatially oriented reasoning. This paper describes an associative visual‐pattern classifier and the automated acquisition of new, spatially oriented reasoning agents that simulate such behavior. They are incorporated into a multi‐agent game‐learning program whose architecture robustly combines agents with conflicting perspectives. When tested on three games, the visual‐pattern classifier learns meaningful patterns, and the pattern‐based, spatially oriented agents generalized from these patterns are generally correct. The accuracy of the contribution of each of the newly created agents to the decision‐making process is measured against an expert opponent, and a perceptron‐Iike algorithm is used to learn game‐specific weights for these agents. Much of the knowledge encapsulated by the new agents was previously inexpressible in the programs representation and in some cases is not readily deducible from the rules.