Using neural networks and Monte Carlo techniques in data science: The value of Google DeepMind in general use

Abstract

Google DeepMind recently unveiled AlphaGo, a body of artificial intelligence (AI) work that purportedly beat world champion Go player, Lee See-Do. It is brilliant insofar as it uses Monte Carlo Search in conjunction with AI to work an optimal path through a near infinitesimal number of moves to triumph over its human adversary. At first glance Monte Carlo Search may seem to be the same as Monte Carlo Simulation but it is far more useful for finding a path to a particular classification. In our example, the classification of interest is the likelihood that a customer will convert to an active customer given a high bid on a digital media advertisement, subject to knowing the potential customer’s environment and journey steps thus far. Using neural networks we are able to produce ‘fiercely accurate’ models that can predict the likelihood of conversion, bringing together enormous amounts of customer journey data, enriched with the marketing expertise native to a business. It is not possible with neural networks, out of the box, to understand why or how the output was formulated. While Monte Carlo Simulation can help unlock some explanatory value in neural networks, the techniques showcased by Google DeepMind bring about a new branch of game theory that can alter the manner in which we approach the discipline of data science in response to business problems.