Ensemble Effort Estimation
Reviewed by Jorge Aranda / 2012-04-17
Kocaguneli2012 Ekrem Kocaguneli, Tim Menzies, and Jacky W. Keung: "On the Value of Ensemble Effort Estimation". IEEE Transactions on Software Engineering, 38(6), 2012, 10.1109/tse.2011.111.
Background: Despite decades of research, there is no consensus on which software effort estimation methods produce the most accurate models.
Aim: Prior work has reported that, given M estimation methods, no single method consistently outperforms all others. Perhaps rather than recommending one estimation method as best, it is wiser to generate estimates from ensembles of multiple estimation methods.
Method: 9 learners were combined with 10 pre-processing options to generate 9 × 10 = 90 solo-methods. These were applied to 20 data sets and evaluated using 7 error measures. This identified the best n (in our case n = 13) solo-methods that showed stable performance across multiple datasets and error measures. The top 2, 4, 8 and 13 solo-methods were then combined to generate 12 multi-methods, which were then compared to the solo-methods.
Results: (i) The top 10 (out of 12) multi-methods significantly out-performed all 90 solo-methods. (ii) The error rates of the multi-methods were significantly less than the solo-methods. (iii) The ranking of the best multi-method was remarkably stable.
Conclusion: While there is no best single effort estimation method, there exist best combinations of such effort estimation methods.
Anybody who has ever done software effort estimation knows that it's a pretty hard thing to do. It's tough even for small individual tasks for someone without practice, and it's a horribly difficult task for large-scale group projects even for estimators with lots of practice. There are many methods that estimators could use, but as Kocaguneli et al remind us, "no single method consistently outperforms all others"—sometimes you're better off using method A, other times, method B would've been more appropriate. Their proposal: to build ensembles of methods, each of them deficient on their own, and to plug them with different automated learners in the hope that these new multi-methods will provide estimates with less error and more consistency.
The multi-methods approach worked well in their (pretty large) dataset of nearly 1,200 projects. This does not mean that the method that came out on top for them will come out on top for you, too. It only means that ensembles of methods are a good workaround for the problem of inconsistency of method efficacy. What the authors propose is for practitioners to learn the basics of machine learning and build method ensembles themselves:
Therefore, our recommendations to practitioners, who are willing to use multi-methods but lack the knowledge of machine learning algorithms are:
- Start with initial 2 learners and build the associated multi-methods
- See the performance of the current multi-methods
- Build new multi-methods only if you are not pleased with the performance of the current ones
That won't be an easy task, but it may be less painful than committing to using a single method that often won't work. If you're interested in doing it, this paper has several references and pointers to get you started.