The emerging role of AI in forecasting
Andrew Tolve surveys the advances in artificial intelligence that can improve the science of forecasting.
By on Feb 16, 2010Andrew Tolve surveys the advances in artificial intelligence that can improve the science of forecasting.
Blockbuster movies like The Terminator and I, Robot would have you believe that machines that can walk, talk, and shoot a gun already live amongst us, or soon will. While the truth is less dramaticeven todays smartest machines cant react, perceive, learn, or deduce anything like a humanartificial intelligence is being used to solve a wide range of challenges for a wide range of industries.
One of these solutions that could benefit pharma is forecasting. Computer scientists have created what are called artificial neural networks to accurately predict future values from a set of input data that, to the human mind, appear noisy and unclear.
Artificial neural networks
Its fairly clear to us all that a brain is considerably more intelligent than a machine, says Laura Baker, AI technical project leader at Software Science, which uses innovative software to help pharmas negotiate marketing challenges. But the brains much more intelligent than a calculator, and we still use them for our everyday math needs because theyre quicker and more reliable.
Artificial neural networks (ANNs) are inspired by biological nervous systems and the way they process information. The human brain contains billions of neurons, and each of those neurons is connected to thousands of other neurons by synapses (essentially, a tube that transfers electrical and chemical impulses). The way the brain learns is by adjusting the strength of these synapses and by adding new connections and deleting old ones.
ANNs are essentially simplified versions of this system. They have neurons (though, due to hardware restrictions, only a fraction of those in the brain) and synaptic weights to connect them.
Forecasting with neural networks
Its all inspired by natural processes around us, says Baker. Its not some mathematician thinking up complex functions in a dark room that no one can understand and relate to. ANNs not only resemble the human brain, they learn like it; that is, by example.
Forecasting is all about predicting a target value from a set of data. With ANNs, computer scientists take a set of input values and ask the machine to predict a target value. They then show the machine the difference between the target value and the value it predicted, have it calculate the difference (the error), and adjust its synaptic weights according to that error. Do that enough times, and you have a trained system that can take input data, no matter how messy, and produce staggeringly accurate forecasts.
Anywhere where the relationships among data are unknown, hard to establish, unclear, noisy, non-linear, thats where you really start to see the benefits of artificial neural networks, says Baker.
Making a better model
ANNs have performed well in studies. The traditional technique for predicting future values in a time series analysis is the autoregressive moving average, or ARMA, model. ARMA is effective for linear data but relatively ineffective for more complex input values. In a study performed in 2006, in which researchers looked at the number of Hepatitis A virus cases from July 2002 to June 2004, the ARMA model produced a 35 percent mean average error rate, compared to a nine percent mean average error rate from the ANN.
Baker warns that even though ANN models have regularly outperformed ARMA models, Its not just a case of out with the old and in with the new. Even the best model can be beaten by a weaker model on occasions, she says. Just like humans. Even the cleverest human [is] not always right no matter how much he thinks he might be.
Thus, computer scientists are using yet another application of artificial intelligence to select which model is most appropriate according to the type of data presented. In other words, if a pharmaceutical company is trying to forecast the number of prescriptions for a drugand to do so, is wading through thousands and thousands of bits of statisticsartificial intelligence can first extract the ten most important bits of that data, then choose which model is most appropriate to analyze it, either the ARMA, the ANN, or a hybrid of the two.
Improving accuracy
By doing this we provide more robust methods and can further enhance and improve on the accuracy that we witnessed in using the models in isolation, says Baker.
Thus, the potential to save money for the pharmaceutical industry is significant. If a pharma forecasts 100,000 future prescriptions for a drug with a traditional ARMA model, that number could be off by 35,000 prescriptions or more (because of the 35 percent mean average error rate). The ANN, on the other hand, would produce a much more manageable error of roughly 9,000 prescriptions. Which means either less wasted money in production or more profit from sales.
Suddenly you can see significant monetary benefit, says Baker. With that in mind, we predict a bright future for forecasting in artificial intelligence.