If you’re deeply involved in the study of artificial intelligence or automated predictive modeling, you may have come across the term “reinforcement learning,” or mapping situations to actions to maximize some type of numerical reward signal. For humans, this process occurs naturally as we grow and experiment with our surroundings and see how our actions influence our rewards.
Reinforcement learning deviates greatly from the normal means by which artificial intelligences are typically programmed. As noted in the book Reinforcement Learning: An Introduction by Richard S. Sutton and Andrew G. Barto, “the most important feature distinguishing reinforcement learning from other types of learning is that it uses training information that evaluates the action taken rather than instructs by giving correct actions.”
In short, reinforcement learning “teaches” machines how to learn from past experience and exploit that information to maximize a reward.
So, what does reinforcement learning have to do with predicting emotions and future behaviors? Quite a lot, actually.
Building on Past Data to Predict Future Outcomes
One of the major issues with most predictive analytical models is that they use a static regression model. In other words, the predictive model is based on a single set of assumptions that are created at the outset, and then do not change to accommodate new information.
Manually adjusting a static regression model requires a human expert working long hours to keep up with the sheer amount of data needed to ensure that the predictive model remains relevant.
By taking advantage of reinforcement learning techniques, machine-learning algorithms can automate the process of learning from past examples much faster than a human worker can integrate instructive examples. This allows for more examples to be integrated into the predictive algorithm, increasing the accuracy of predictions.
The Importance of Emotions in Customer Conversations
For example, say that you have a predictive voice analytics program for call centers that studies human speech patterns to determine emotional tone and behavior, then predict likely future behaviors.
Here, the “reward” for the algorithm could be defined as a positive call outcome where the customer agrees to pay for something, such as a product, service, or outstanding debt. Rather than taking action itself to gain a reward, the algorithm dissects how the customer speaks, breaking specific vocal features into feature vectors that can be easily categorized by the program.
Determining the customer’s emotional behavior and tone allows the algorithm to create a predictive analysis of what the customer’s future behaviors will be. This allows the algorithm to rank customer interactions by likelihood of a positive outcome on a follow-up call.
The Benefits of Algorithms That Can Crunch Through Massive Data
Over the course of thousands of calls, an algorithm that uses reinforcement learning will collect data on call results both positive and negative, and link individual feature vectors (emotional cues) from each call to the end result. Aggregating results data allows the algorithm to increase the accuracy of its predictive model based on actual past results.
Reinforcement learning helps to make a predictive analytics program better at making predictions of customer behavior based on past experience, and saves you the trouble of hiring and dedicating a human expert to the task of making use of the data.