Abstract
Cost-benefit analysis (CBA) is a systematic approach for calculating and analyzing the costs of a project. Soft computing approaches are increasingly applicable to deal with the inherent uncertainties in CBA. In this paper, a Mamdani fuzzy system has been developed for cost-benefit analysis, optimized using genetic algorithms. The model's performance is evaluated based on both interpretability and accuracy features.
Introduction
Cost-Benefit Analysis is a cornerstone of project planning, yet it is often plagued by uncertainty and imprecision. Traditional analytical models frequently struggle to account for the "gray areas" of project variables, such as fluctuating sales quantities or varying overhead costs. While soft computing approaches like fuzzy logic are well-suited for this kind of ambiguity, they often face a classic engineering dilemma: the Interpretability-Accuracy trade-off.
A model can be highly accurate but act as an incomprehensible "black box," or it can be simple and transparent but lack precision. This research presents a Genetic Fuzzy System (GFS) designed to automate and optimize CBA for project management. By combining the human-like reasoning of Mamdani fuzzy logic with the optimization power of Genetic Algorithms (GA), the authors developed a model that achieves 98% accuracy in profit estimation while remaining highly interpretable for human decision-makers.
Methodology
The researchers utilized a Genetic Fuzzy Rule-Based System (GFRBS). The methodology integrates two primary components to balance performance and clarity.
Mamdani Fuzzy System
Unlike standard logic that uses binary (true/false) values, this system uses linguistic variables (e.g., "Poor," "Average," "High") to categorize inputs. This allows the model to process data in a way that mirrors human expert judgment, mapping raw currency values to intuitive levels. For example, a "Cost on Sale" between 1800–2050 INR might be categorized as "Poor." This abstraction allows the model to handle "noisy" data more gracefully than rigid mathematical formulas.
Genetic Algorithm (GA) Optimization
To refine the system, a Genetic Algorithm was employed to optimize rule selection. The GA "evolves" the best set of rules over 5,000 generations, using processes like crossover and mutation to prune redundant logic and identify the most effective rule combinations. This automated pruning prevents "rule explosion," ensuring the system remains manageable.
The implementation was performed using GUAJE (Generating Understandable and Accurate fuzzy models in a Java Environment), an open-source tool specifically designed to balance the competing needs of accuracy and interpretability.
Data and Experimental Setup
The paper utilizes a dataset consisting of 10 primary data points/rules. Each entry includes four input variables:
- Cost on Sale (COS)
- Quantity of Sale (QOS)
- Cost at Variation (CAV)
- Cost at Fixed (CAF)
The target output is Profit Gained (PG). The Genetic Algorithm successfully converged over 5,000 generations with a population length of 50, a crossover probability of 0.8, and a mutation probability of 0.1.
Key Results and Discussion
The proposed GFRBS model demonstrated impressive performance metrics, successfully bridging the gap between accuracy and transparency:
- High Accuracy: The model achieved a 98% accuracy rate, indicating a very high correlation between the model's profit estimates and the actual data.
- Quantifiable Interpretability: The paper uses formal metrics like Nauck’s Index and an Interpretability Index (calculated at 0.115) to provide a mathematical basis for how "readable" the model is to a human observer.
- Optimized Complexity: The final model settled on just 10 rules with an average rule length of 4, suggesting a lean and efficient logic structure.
Conclusion and Future Work
The GFRBS model provides a robust framework for CBA, offering high precision without sacrificing the "why" behind its predictions. However, a primary limitation is the current scale of the dataset. Future research should explore how the model scales to massive, multi-variable industrial projects. Additionally, the authors express interest in exploring Interval Type-2 Fuzzy Systems, which are even more robust at modeling "uncertainty about uncertainty," potentially further improving reliability in highly volatile markets.
Summary updated