Summary
The case concerns the development of a flavor prediction innovation (neural networks) by a leading British brewing company, Coors Brewers Ltd. Although a large amount of chemical and sensory data is available, techniques for determining the relationship between the two variables are lacking. Neural networks offer a quicker prediction method than the test panels traditionally used to test beer flavors. Coors, motivated by the need to expand its market presence, set out to find a technique that would determine beer flavor based exclusively on analytical (chemical) data. The technique would enable the firm to create brands that could meet the diverse tastes and expectations of drinkers.
Coors developed the technique in a three-phase project. Initially, Coors used a single neural network to determine how chemical composition influences flavor based on sensory and analytical data it had collected. To implement it, Coors trained the MLP neural network using different combinations of sensory and analytical data.
This enabled the firm to analyze each quality and sensory output individually. Normalized data allowed the network to compare different outputs and minimize network error. However, the technique could not define meaningful relationships between output and input data because only a single quality was analyzed at a time. This limited data variability. In addition, ‘noise’ created by extraneous inputs affected the technique’s effectiveness.
Coors developed an improved version that included a software switch to eliminate the effect of insignificant inputs and reduce the network error. The method was exhaustive as it evaluated all input combinations. However, it yielded a large dataset (combinations) that could not be solved with the available computational methods. To overcome this limitation, Coors developed a genetic algorithm to determine the input/output combination that could yield a lower network error and thus, predict flavor more accurately. The results indicated that the trained genetic algorithm could accurately predict a number of flavors using trained chemical data. The current technique can only predict a few flavors. Moreover, it does not take into consideration other sensory factors.
Why is beer flavor important to Coors’ profitability?
Customer beer preferences and choices are never constant. A customer’s drink choice often depends on the occasion, settings, and psychological state. Coors, as one of the leading firms in the British brewing industry, intends to expand its flavors to reflect the changing and diverse customer preferences and needs. The beer flavor is one way the company can differentiate its products in order to provide customers a broad array of drinks that suit all situations.
The beer flavor is an important quality that drinkers take into consideration in choosing a brand-appropriate for a particular occasion. The traditional testing methods (panel tests) are time-consuming. In contrast, alternative techniques that are faster and accurate can give a company a strong competitive advantage in the industry. Coors aims to come up with a novel method of predicting flavor based on chemical data alone. This will enable the company to develop beer brands that meet customer expectations.
What is the objective of the neural network used at Coors?
The aim of Coors’ neural network is to select input/output combinations with the least network error in order to predict beer flavors more accurately. Such combinations will improve the accuracy and speed of predicting beer flavors. The alternative method (panel testing) is slow and requires multiple inputs. In contrast, neural networks have the ability to predict beer flavor using only the chemical input data.
Coors implemented different versions of neural networks in a bid to develop a refined prediction technique. The genetic algorithm technique developed synthesizes chemical inputs and releases sensory outputs, which define a beer’s flavor. The company uses input and output data that have been gathered from panel testing over the years. The role of the neural network is to model the link between inputs and outputs. Coors has improved the performance of its neural networks over the years producing a highly accurate prediction tool. The prediction accuracy of the genetic algorithm tool stems from its ability to reduce network error.
Why were the results of Coors’ neural network initially poor, and what was done to improve the results?
Coors launched its project by implementing a single neural network. This product was made up of a two-layered MLP sourced from an external developer. It focused on a single input (chemical quality) and output (flavor). Using normalized data drawn from input combinations, the network was trained to do cross-comparisons of different sensory outputs. However, the single neural network had two significant limitations that affected the quality of the results. First, it used a single input variable, which reduced data variability. This meant that no meaningful relationships could be modeled. Second, ‘noise’ caused by extraneous inputs affected the network’s effectiveness.
To improve data variability, Coors expanded the product range. This generated more analytical input data for training the network. The second modification was the introduction of a software switch that allowed the training of the network using input/output combinations making up the probability space. This exhaustive search, though effective in removing the ‘noise’, generated many combinations. The huge number of combinations per each sensory attribute (up to 16.7 per flavor) was mathematically impossible to compute. A gene algorithm method, which could select relevant inputs, was developed to replace this method. The gene algorithm technique is a more accurate method of searching inputs with minimal network error for accurate flavor prediction.
What benefits might Coors derive if this project is successful?
Coors will gain many benefits from the ongoing project. First, the technology will help Coors in product differentiation. This will ensure that it offers unique and more attractive products than those offered by its competitors. An improved version of this technique will have the capacity to predict more flavors. The technique will enable the company to select chemical inputs that produce unique flavors.
Coors will also be able to produce a broad range of flavors that reflect diverse customer tastes and expectations. This will increase the firm’s market share and give it a strong competitive advantage over its rivals in the beer brewing industry. Since it owns the technology, Coors will also benefit from the exclusive rights that come with patenting an innovation. It can sell or lease the technology to other brewers for a fee.
What modifications would you make to improve the results of beer flavor prediction?
The current technology can only predict a small range of flavors. Moreover, it leaves out certain compounds that influence beer flavor. To enhance prediction, I will analyze all the compounds that affect flavor to enrich the available chemical data. With a broad array of input data, more flavors can be determined accurately. Moreover, I will factor in the effect of physical variables related to drinking as they affect flavor sensation. These include ‘mouth-feel’ variables and the appearance of the beer. I will develop a sensory profile of a typical drinker to aid in the prediction process. To improve the neural network’s prediction accuracy, I will include the physical variables in the input data fed into the program.
Reference List
Sharda R, Delen D, Turban E. Business Intelligence and Analytics: Systems for Decision Support. Upper Saddle River, NJ: Prentice-Hall; 2014.