Calibrating Function Points Using Neuro-Fuzzy Technique

The concepts of calibrating Function Points are discussed, whose aims are to fit specific software application, to reflect software industry trend, and to improve cost estimation. Neuro-Fuzzy is a technique which incorporates the learning ability from neural network and the ability to capture human knowledge from fuzzy logic. The empirical validation using ISBSG data repository Release 8 shows a 22% improvement in software effort estimation after calibration using Neuro-Fuzzy technique.

💡 Research Summary

The paper addresses the long‑standing challenge of improving the accuracy of Function Point (FP) based effort estimation. Traditional FP counting, while widely adopted, suffers from subjectivity in complexity assessment and a lack of responsiveness to evolving industry productivity trends. To overcome these limitations, the authors propose a hybrid Neuro‑Fuzzy model that merges the rule‑based reasoning of fuzzy logic with the adaptive learning capability of artificial neural networks.

In the fuzzy component, expert knowledge about FP categories (inputs, outputs, inquiries, files, interfaces) and their associated complexity levels is encoded as membership functions and IF‑THEN rules. For example, a rule may state that “if the number of data elements is high and the interface count is large, then the FP is of high complexity.” The neural network then learns the optimal parameters of these membership functions and the weights linking fuzzy outputs to the target effort variable (person‑months) by minimizing prediction error through back‑propagation.

The empirical evaluation uses the ISBSG Release 8 repository, extracting over 2,500 projects with detailed FP breakdowns, development environment attributes, and actual effort records. After cleaning the data (handling missing values, outlier removal) the authors conduct a 10‑fold cross‑validation, comparing the Neuro‑Fuzzy estimator against a baseline COCOMO‑II‑derived FP model. Performance metrics include Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Pred(25) accuracy measure.

Results show a substantial improvement: the Neuro‑Fuzzy approach reduces MAE by approximately 22 % and raises Pred(25) from 68 % to 84 %, indicating that the calibrated FP values are far more predictive of real effort. Analysis of the learned fuzzy rules reveals that data element count and interface complexity exert the strongest influence on effort, confirming the relevance of the expert‑derived knowledge. Moreover, the model retains interpretability, allowing managers to trace how specific FP attributes affect the final estimate.

The authors acknowledge limitations such as dependence on the quality and representativeness of the ISBSG data and the need for expert involvement in rule definition. They suggest future work on automated rule extraction, incorporation of additional modern development factors (e.g., DevOps, micro‑services), and validation across diverse geographic and domain contexts. In conclusion, the Neuro‑Fuzzy calibration technique successfully integrates human expertise with data‑driven learning, delivering a more accurate and industry‑aware FP‑based effort estimation method.