Testing Environmental Models supported by Machine Learning

• Valdivino Alexandre de Santiago INPE
• Leoni Augusto Romain da Silva INPE
• Pedro Ribeiro de Andrade Neto INPE

Resumo

In this paper we present a new methodology, DaOBML, to test environmental models whose outputs are complex artifacts such as images (maps) or plots. Our approach suggests several test data generation techniques (Combinatorial Interaction Testing, Model-Based Testing, Random Testing) and digital image processing methods to drive the creation of Knowledge Bases (KBs). Considering such KBs and Machine Learning (ML) algorithms, a test oracle assigns the verdicts of new test data. Our methodology is supported by a tool and we applied it to models developed via the TerraME product. A controlled experiment was carried out and we conclude that Random Testing is the most feasible test data generation approach for developing the KBs, Artificial Neural Networks present the best performance out of six ML algorithms, and the larger the KB, in terms of size, the better.

Palavras-chave: Combinatorial Interaction Testing, Digital Image Processing, Empirical Software Engineering, Environmental Modeling, Machine Learning, Model-Based Testing, Random Testing