An Intelligent Approach to Software Cost Prediction

Good software cost prediction is important for effective project management such as budgeting, project planning and control. In this paper, we present an intelligent approach to software cost prediction. By integrating the neuro-fuzzy technique with the well-accepted COCOMO model, our approach can make the best use of both expert knowledge and historical project data. Its major advantages include learning ability, good interpretability, and robustness to imprecise and uncertain inputs. The validation using industry project data shows that the model greatly improves prediction accuracy in comparison with the COCOMO model.

💡 Research Summary

The paper addresses the persistent challenge of accurately forecasting software development costs, a critical factor for budgeting, scheduling, and risk management. While the Constructive Cost Model (COCOMO) has long been the industry standard, its reliance on static coefficients and crisp input values limits its adaptability to the nuanced, often ambiguous realities of modern projects. To overcome these shortcomings, the authors propose a hybrid intelligent model that integrates a neuro‑fuzzy system—specifically an Adaptive‑Network‑Based Fuzzy Inference System (ANFIS)—with the traditional COCOMO framework.

The methodology begins by mapping COCOMO’s fifteen cost drivers and the size metric (KLOC) into fuzzy linguistic variables such as “low,” “medium,” and “high.” Each linguistic term is associated with a membership function (triangular or Gaussian), allowing partial degrees of truth for inputs that do not fit neatly into discrete categories. Expert knowledge is used to construct an initial rule base that mirrors COCOMO’s empirical relationships, thereby preserving interpretability. The ANFIS architecture then treats these fuzzy rules as a layered neural network, enabling simultaneous learning of membership‑function parameters and rule weights through a hybrid training algorithm that combines forward fuzzy inference with backward error propagation.

For empirical validation, the authors collected a dataset of 120 real‑world projects from a major Korean IT firm, encompassing size, driver ratings, and actual effort/cost outcomes. The dataset was randomly split into 70 % training and 30 % testing subsets. Model performance was assessed using two widely accepted metrics: Mean Magnitude of Relative Error (MMRE) and the percentage of predictions within 25 % of actual values (Pred(25)). The neuro‑fuzzy‑augmented COCOMO achieved an MMRE of 0.18 versus 0.23 for the baseline COCOMO—a 22 % reduction—and a Pred(25) of 68 % compared with 53 % for the traditional model, indicating a 15 percentage‑point improvement. Notably, the fuzzy preprocessing step conferred robustness to missing or imprecise driver inputs, a common issue in practice.

The authors discuss several strengths of their approach: (1) adaptive learning that captures evolving project characteristics, (2) retained transparency through fuzzy rules that can be inspected by project managers, and (3) resilience to uncertain or incomplete data. They also acknowledge limitations, such as the initial dependence on expert‑defined rules and increased computational overhead during training on larger datasets. Future research directions include automated rule extraction, deeper neural architectures for multi‑objective optimization (balancing cost, schedule, and quality), and extensive cross‑industry testing.

In conclusion, the study demonstrates that embedding neuro‑fuzzy intelligence within the COCOMO paradigm yields a more accurate, interpretable, and robust cost prediction tool, bridging the gap between empirical models and data‑driven learning. The empirical results provide compelling evidence that the hybrid model can be deployed in real‑world settings to enhance project planning and control.