Reducing Biases in Individual Software Effort Estimations: A Combining Approach

In this section, we first differentiate the “composite estimation
method” term frequently used in the software effort estimation field
from our mentioned “combining estimation method” in this paper,
and then we examine current combining methods or practices
employed in software effort estimation, discuss their drawbacks, and
propose the OLC method at last.

Based on reviews of software effort estimation approaches[4-6],
three current typical “composite estimation methods” are frequently
referred in the software effort estimation field, they are Cost
Estimation Benchmarking and Risk Analysis (COBRA)[3],
Incorporating Bayesian analysis to improve the accuracy of
COCOMO II[2, 31] and Analogy-Based tools[32, 33]. However, the
three methods are different from J.Scott Armstrong’s definition of
combining forecasts. Technically, they all combine experiential
approaches with data-driven modeling, but expert knowledge used
in the three methods is just a part of the modeling process: COBRA
uses expert knowledge to construct the causal model; Analogy-
Based tools use it to define similarity functions, and COCOMO II
combine it with historical data to calibrate the value of cost drivers
by Bayesian Analysis.

This kind of composite technique is
“internal” combining and in some sense, they all belong to
individual techniques. In this paper, we refer “combining estimation
method” as to combine the outputs (final effort estimates) of
individual methods by simple average or a sophisticated weighting
algorithm, as Armstrong’s definition mentioned in section 2.1. For
differentiating from “internal” combining, we call it “external”
combining.