Reference no: EM13332817

Today's customer has a wide variety of options regarding what he or she drinks. A drinker's choice depends on various factors, including mood, venue, and occasion. Coors' goal is to ensure that the customer chooses Coors brands no matter what the circumstances are.

According to Coors, creativity is the key to long-term success. To be the customer's choice brand, Coors needs to be creative and anticipate the customer's ever so rapidly changing moods. An important issue with beers is flavor; each beer has a distinctive flavor. These flavors are mostly determined through panel tests. However, such tests take time. If Coors could understand the beer flavor based solely on its chemical composition, it would open up new avenues to create beer that would suit customer expectations.

The relationship between chemical analysis and beer flavor is not clearly understood yet. Substantial data exist on the chemical composition of a beer and sensory analysis. Coors needed a mechanism to link those two together. Neural networks were applied to create the link between chemical composition and sensory analysis.

Over the years, Coors Brewers Ltd. Has accumulated a significant amount of data related to the final product analysis, which has been supplemented by sensory data provided by the trained in-house testing panel. Some of the analytical inputs and sensory outputs are shown in the following table:

Analytical Data: Inputs Sensory Data: Outputs
Alcohol Alcohol
Color Estery
Calculated bitterness Malty
Ethyl acetate Grainy
Isobutyl acetate Burnt
Ethyl butyrate Hoppy
Isoamyl acetate Toffee
Ethyl hexanoate Sweet

A single neural network, restricted to a single quality and flavor, was first used to model the relationship between the analytical and sensory data. The neural network was based on a package solution supplied by NeuroDimension,Inc. The neural network consisted of an MLP architecture with two hidden layers. Data were normalized within the network, thereby enabling comparison between the results for the various sensory outputs. The neural network was trained (to learn the relationship between inputs and outputs) through the presentation of many combinations of relevant input/output combinations. When there was no observed improvement in the network error in the last 100 epochs, training was automatically terminated. Training was carried out 50 times to ensure that a considerable mean network error could be calculated for comparison purposes. Prior to each training run, a different training and cross-validation dataset was presented by randomizing the source data records, thereby removing any bias.

This technique produced poor results, due to two major factors. First, concentrating on a single product's quality meant that the variation in the data was pretty low. The neural network could not extract useful relationships from the data. Second, it was probable that only one subset of the provided inputs would have an impact on the selected beer flavor. Performance of the neural network was affected by "noise" created by inputs that had no impact on flavor.

A more diverse product range was included in the training range to address the first factor. It was more challenging to identify the most important analytical inputs. This challenge was addressed by using a software switch that enabled the neural network to be trained on all possible combinations of inputs. The switch was not used to disable a significant input; if the significant input were disabled, we could expect the network error to increase. If the disabled input was insignificant, then the network error would either remain unchanged or be reduced due to the removal of noise. This approach is called an exhaustive search because all possible combinations are evaluated. The technique, although conceptually simple, was computationally impractical with the numerous inputs; the number of possible combinations was 16.7 million per flavor.

A more efficient method of searching for the relevant inputs was required. A genetic algorithm was the solution to the problem. A genetic algorithm was able to manipulate the different input switches in response to the error term from the neural network. The objective of the genetic algorithm was to minimize the network error term.

When this minimum was reached, the switch settings would identify the analytical inputs that were most likely to predict the flavor.