This paper proposes that the generalisation capabilities of a case-based reasoning system can be evaluated by comparison with a `rote-learning' algorithm which uses a very simple generalisation strategy. Two such algorithms are deened, and expressions for their classiication accuracy are derived as a function of the size of training sample. A series of experiments using artiicial and`natural' data sets is described in which the learning curve for a case-based learner is compared with those for the apparently trivial rote-learning learning algorithms. The results show that in a number of`plausible' situations, the learning curves for a simple case-based learner and thèmajority' rote-learner can barely be distinguished, although a domain is demonstrated where favourable performance from the case-based learner is observed. This suggests that the maxim of case-based reasoning that`similar problems have similar solutions' may be useful as the basis of a generalisation strategy only in selected domains.

In order to learn more about the behaviour of case-based reasoners as learning systems, we form-alise a simple case-based learner as a PAC learning algorithm, using the case-based representation hCB; i. We rst consider a `naive' case-based learning algorithm CB1(H) which learns by collecting all available cases into the case-base and which calculates similarity by counting the number of features on which two problem descriptions agree. We present results concerning the consistency of this learning algorithm and give some partial results regarding its sample complexity. We are able to characterise CB1(H) as a `weak but general' learning algorithm. We then consider how the sample complexity of case-based learning can be reduced for speciic classes of target concept by the application of inductive bias, or prior knowledge of the class of target concepts. Following recent work demonstrating how case-based learning can be improved by choosing a similarity measure appropriate to the concept being learnt, we deene a second case-based learningàlgorithm' CB2 which learns using the best possible similarity measure that might be inferred for the chosen target concept. While CB2 is not an executable learning strategy (since the chosen similarity measure is deened in terms of a priori knowledge of the actual target concept) it allows us to assess in the limit the maximum possible contribution of this approach to case-based learning. Also, in addition to illustrating the role of inductive bias, the deenition of CB2 simpliies the general problem of establishing which functions might be represented in the form hCB; i. Reasoning about the case-based representation in this special case has therefore been a little more straightforward than in the general case of CB1(H), allowing more substantial results regarding representable functions and sample complexity to be presented for CB2. In assessing these results, we are forced to conclude that case-based learning is not the best approach to learning the chosen concept space (the space of monomial functions). We discuss, however, how our study has demonstrated, in the context of case-based learning, the operation of concepts well known in machine learning such as inductive bias and the trade-oo between computational complexity and sample complexity.

Hull, J. H., and Sarubbi, F. A., Jr (1976). Gentamicin serum concentrations: pharmacokinetic predictions. Annals of Internal Medicine, 85, 183-189. Kaye, D., Levison, M. E., and Labovitz, E. D. (1974). The unpredictability of serum concentrations of gentamicin: pharmacokinetics of gentamicin in patients with normal and abnormal renal function. Journal ofInfectious Diseases, 130, 150-154. Reeves, D. S., and Bywater, M. J. (1975). Quality control of serum gentamicin assays-experience of national surveys. Journal of Antimicrobial Agents and Chemotherapy, 1, 103-116. Schwartz, G. J., Haycock, G. B., Edelmann, C. M., Jr, and Spitzer, A. (1976). A simple estimate of glomerular filtration rate in children derived from body length and plasma creatinine. Pediatrics, 58, 259-263.

During a screening programme for the detection of CF using the meconium albumin technique, the overall false-positive rate was found to be approximately 1%. When the gestational age of the infants was taken into account the false-positive rate was found to be significantly higher in preterm (8%) as compared to term infants (0.55%). This was due largely but not solely to the presence of occult blood. Possible explanations for these findings are discussed and attention drawn to the limitation of meconium albumin content as a screening technique for CF in preterm infants.

A simple immunochemical technique utilizing single radial immunodiffusion for the demonstration of albumin in human meconium is described. A comparison of this technique with the more widely used sulphosalicylic acid precipitation method is reported. We also have attempted to determine the normal levels of albumin in meconium.

about the trial so that we will be notified of any children who are found to have CF who were born in either of the hospitals since the beginning of the trial. In this way we hope to find out how many cases of CF have been missed by the test. As far as we know this is the first attempt at a prospective screening programme for CF using this technique. The time taken for the test is 10 to 15 seconds and the cost of materials is just under 2p per test. In one of the hospitals the paediatric Senior House Officer does the tests while in the other the midwives and paediatric nursing staff do them. We hope that this limited trial will stimulate other maternity units to carry out similar surveys so that the efficiency of the method can be more rapidly assessed. It is of interest that if we had started the trial four weeks earlier we would also have detected a case of intestinal lymphangiectasia which presented at the age of 3 weeks with gross oedema and a protein-losing enteropathy.