Neural Nets Raise the Roof

have problems weighing the various factors that influence the need for roof repair. So, how can you do away with these traditional, bulky, tools, while improving service and the ability to train new staff? We devised a PDA-based expert system that uses artificial neural networks to provide a roofing advisor. Roofers might be familiar with a large number of roof types but concentrate on those for which maintenance is practical. We identified the prime inputs for each of these roof types along with the expected results. For example, for a concrete tiled roof (see Figure 1a), the amount of fungal growth and an estimation of wear relative to the expected lifespan would be suitable measures. We identified these from all possible inputs through principle component analysis. We then created test cases, which let an expert roofer and us agree on system results given particular input values. Because each roof type has different properties, the number and type of input parameters vary. For example, a concrete tile roof has two input values (fungal growth and wear), while a pressed metal tile roof has five similar but distinct input values (fungal growth, chip loss, base coat wear, rust in pans, and rust in valleys). A roof's location and profile add to the complexity. Location is a physical location (usually " suburb, " although it could be a street address). Profile is the shape of the tile or roof material. For instance, a concrete tile roof's profile might be Atlas, Concurve, Mid-Petrous, or any one of hundreds of other shapes. Although we initially used default values, we included adjustments to add realism. For example, a roof in Invercargill will last about 40 years or less, while the same roof in Timaru might last 55 years. With a range of real examples, artificial neural networks can later learn these factors. The input values are continuous and are treated as such by the user interface and within the decision-making process. Discrete values are hidden from the user—the mental model and display are a continuum. Conversions from discrete to continuous values and vice versa occur in the application's Translation layer (see the sidebar). However, existing paper-based forms for assessing roofs describe the values in terms of good, moderate, and bad conditions. These " fuzzy values " suggested the use of fuzzy values in a FuNN (fuzzy neural network, a variation of the multilayer perceptron). A FuNN's …