Raúl E. Valdés-Pérez

Principles of human¿computer collaboration for knowledge discovery in science

An important problem in computational scientific discovery is to identify, among the diversity of discovery programs written in various sciences, a commonality that will take a next step beyond the acknowledged general-but weak-framework of heuristic search. We characterize discovery in science as the generation of novel, interesting, plausible, and intelligible knowledge about the objects of study. We then analyze four current machine discovery programs in chemistry, medicine, mathematics, and linguistics according to how their design, or the circumstances of their application, heighten the chances of finding knowledge that has all four properties. Some general patterns emerge, although some strategies seem idiosyncratic. Our candidate for a commonality, which focuses on human factors, can be used pragmatically to evaluate and compare the designs of discovery programs that are intended to be used as collaborators by scientists.

Conjecturing hidden entities by means of simplicity and conservation laws: machine discovery in chemistry

Abstract We show that combinatorial search, constrained by experimental evidence, domain knowledge, and simplicity, is sufficient to discover credible explanatory hypotheses in a scientific task of current importance. The MECHEM system automates the scientific task of determining the pathway of steps underlying a chemical reaction. MECHEM carries out a systematic search for pathway hypotheses that satisfy constraints and explain the available evidence. During its search, the system conjectures unseen entities as needed to formulate hypotheses, but only after determining by experimentation or other means that all simpler hypotheses are inadequate. The generality of MECHEMs methods is established by pointing out similarities to other scientific tasks and programs.

Online measures of basic language skills in children with early focal brain lesions

Twenty children with early focal lesions were compared with 150 age-matched control subjects on 11 online measures of the basic skills underlying language processing, a digit span task, and 6 standardized measures. Although most of the children with brain injury scored within the normal range on the majority of the tasks, they also had a disproportionately high number of outlier scores on the reaction time tests. This evidence for a moderate impairment of the basic skills underlying language processing contrasts with other evidence suggesting that these children acquire normal control of the functional use of language. Furthermore, these children scored within the normal range on a measure of general cognitive ability, suggesting that there is no particular sparing of linguistic functions at the expense of general cognitive functions. Using the MPD procedure (Valdés-Pérez & Pericliev, 1997), we found that the controls and the five clinical groups could be best distinguished with two measures of online processing (word repetition and visual number naming) and one standardized test subcomponent (the CELF Oral Directions subtest). The 12 children with left hemisphere lesions scored significantly lower than the 8 other children on the CELF-RS measure. Within the group of children with cerebral infarct, the nature of the processing disability could be linked fairly well to site of lesion. Otherwise, there was little relation between site or size of lesion and the pattern of deficit. These results support a model in which damage to the complex functional circuits subserving language leads to only minor deficits in process efficiency, because of the plasticity of developmental processes.

Machine discovery in chemistry: new results

Abstract Earlier we proposed an idea for conjecturing unseen entities in science, and described its application within MECHEM to the chemistry task of inferring the mechanism of a chemical reaction based on experimental evidence. However, the program was a prototype, and lacked several capabilities that rendered it incompetent on current science. We now describe extensions that enable reasoning about the molecular structural transformations that are the focus of modern chemistry. We also report successful applications of MECHEM to chemical problems of current interest, and point out subsequent machine discovery work that the MECHEM project has strongly influenced. These new results demonstrate the efficacy and generality of the original idea for machine discovery, and vindicate the research strategy of emphasizing specific task competence and deferring concerns with generality.

Scientific discovery and simplicity of method

3-Benzoyl-3-thiocyanatopropionic acid, alkyl esters are prepared by the reaction of a 3-benzoyl-3-halopropionic acid alkyl ester with an alkali metal thiocyanate. The products so produced exhibit antitubercular activity.

Human/computer interactive elucidation of reaction mechanisms: application to catalyzed hydrogenolysis of ethane

The computer program MECHEM is a tool for interactive elucidation of reaction mechanisms. Such programs confront severe combinatorial problems on complex multistep reactions, hence the practicality of these efforts is doubtful. We report the first evidence that MECHEM is successfully applicable to practical experimental problems, in this case, catalyzed hydrogenolysis of ethane.

Demonstration of hierarchical document clustering of digital library retrieval results

As digital libraries grow in size, querying their contents will become as frustrating as querying the web is now. One remedy is to hierarchically cluster the results that are returned by searching a digital library. We demonstrate the clustering of search results from Carnegie Mellons Informedia database, a large video library that supports indexing and retrieval with automatically generated descriptors.

Some Recent Human-Computer Discoveries in Science and What Accounts for Them

My collaborators and I have recently reported in domain science journals several human-computer discoveries in biology, chemistry, and physics. One might ask what accounts for these findings, for example, whether they share a common pattern. My conclusion is that each finding involves a new representation of the scientific task: The problem spaces searched were unlike previous task problem spaces. Such new representations need not be wholly new to the history of science; rather, they can draw on useful representational pieces from elsewhere in natural or computer science. This account contrasts with earlier explanations of machine discovery based on the expert system view. My analysis also suggests a broader potential role for (AI) computer scientists in the practice of natural science.

Oxidative carbonylation of phenylacetylene catalyzed by Pd(II) and Cu(I): Experimental tests of forty-one computer-generated mechanistic hypotheses

Abstract We describe an experimental study of the reaction mechanism of phenylacetylene oxidative carbonylation to methyl ester of phenylpropiolic acid catalyzed by Pd(II) and Cu(I), PhCCH+CO+MeOH+2NaOAc+2CuCl 2 →PhCCCOOMe+2AcOH+2NaCl+2CuCl, which was closely guided by recent computational research on the generation of reaction mechanisms. Our initial mechanistic studies of this reaction were based on informal (non-computer-generated) mechanistic hypotheses. When experiments at 20°C and 1 atm led us to reject four of five mechanistic possibilities for the reaction, we turned to formulating new hypotheses with the aid of the computer programs ChemNet, which generated a reaction network consisting of 233 elementary steps, and MECHEM, which uncovered 41 simplest hypothetical pathways from within the reaction network. Our subsequent analysis of these 41 hypothetical mechanisms suggested a highly informative experiment based on the CH 3 OH/CH 3 OD kinetic isotope effect. The ratio between the rates of ester formation in nondeuterated and deuterated methanol was close to unity, suggesting that O–H bond scission occurs after the rate-limiting transmetalation step CuCCPh+PdCl 2 →ClPdCCPh+CuCl. This experiment led to rejecting 32 out of the 41 hypotheses. Four more mechanisms were rejected based on the results of preliminary experimental studies. Further work is needed to discriminate among the five remaining mechanisms.

Algebraic Reasoning about Reactions: Discovery of Conserved Properties in Particle Physics

Kocabas (1991) describes a situation from particle physics in which quantum properties and conservation laws are postulated from lists of observed and unobserved reactions. Kocabas also presents a program named BR-3 that can rediscover some accepted quantum properties from textbook data, although it fails on a more difficult example from the same source. This paper describes PAULI, a program that solves the same task as BR-3 but uses a different problem-solving model. PAULI produces different, simpler solutions than does BR-3, and it can also handle the problematic example. After comparing the two programs, we conclude that PAULI offers distinct advantages over its predecessor, which we attribute to an algebraic approach to reasoning about sets of reactions.