This paper discusses a case study of the application of neural networks to the evaluation of classroom instruction in Logo programming. In experiment 1, the subjects were ninety-six junior high school students in third grade. Their reflections of thinking processes were measured with the RSTC (Reflection Scale of Thinking Processes on Computer Programming : MORIYAMA, et al. 1996). Students' programming abilities were also measured by programming tests after the instruction. We selected 5 out of 24 items from the RSTC, which have significant correlation with programming test scores. The answers to these five items were used for acquiring data on back propagation neural networks with bias units. As a result of learning simulation, the correlation between measured and predicted scores obtained was 0.821. The mean score of errors was 0.014. In experiment 2, the subjects were thirty-five junior high school students in third grade. Their reflections were measured with a modified RSTC (5 items), and their programming abilities were also evaluated with debug and coding tests. As a result of recognizing simulation by using neural networks which was generated in experiment 1, the correlation between measured and predicted scores was 0.631. The absolute value of errors was under 0.1 in 89.1% of the total of predicted scores. From these results, it was suggested that a predictive evaluation system of learning effects in instruction on Logo programming could be constructed by using neural networks.
View full abstract