On cars, old, new and future; science & technology; vintage airplanes, computer flight simulation of them; Sherlockiana; our English language; travel; and other stuff
HOW DO WE make a decision? To perform a particular action. To buy a product. To elect a person. Predictions and analyses of decision-making are research areas in both the social sciences and in engineering.
Two articles in AAAS Science magazine describe a mixture of predictive theories enhanced by machine-generated analyses of large datasets. The technical paper is “Using Large-Scale Experiments and Machine Learning to Discover Theories of Human Decision-Making,” by Joshua C. Peterson, et al., Science, June 11, 2021. “Machine-generated Theories of Human Decision-making,” by Sudeep Bhatia and Lisheng He, is a summary of the full paper presented in the same issue. Here are tidbits gleaned from both.
Background. As noted by Peterson and colleagues, “The use of large datasets has revolutionized machine learning, computer vision, and artificial intelligence. Our study is one of the first to use a similar methodology in systematically investigating theories of human cognition.”
Applying This Greater Computer Power. Bhatia and He observe, “Instead of relying on the intuitions and (potentially limited) intellect of human researchers, the task of theory generation can be outsourced to powerful machine-learning algorithms…. The flexibility of deep networks allows them to find better mathematical implementations for these properties and learn other properties necessary for describing data that have not been previously identified by human researchers.”
Large-scale datasets. Peterson and colleagues write, “Historically, datasets on risky choice [i.e., gambles with differing probabilities and payoffs] have been small…. To address these challenges, we collected a large dataset of human decisions for almost 10,000 choice problems presented in a format that has been used in previous evaluations of models of decision-making. This dataset includes >30 times the number of problems in the largest previous dataset.”
You may recall that machine learning thrives on such large datasets. In a sense, it sifts through the data to identify patterns from which it makes inferences.
The Mathematics of It All. The researchers continue, “We then used this dataset to evaluate differentiable decision theories that exploit the flexibility of deep neural networks but use psychologically meaningful constraints to pick out a smooth, searchable language of candidate theories with shared assumptions.”
Here, A and B are pairs of gambles or “choice problems.” P(A) identifies the probability of a decision maker choosing A after consideration of both choices.
Conclusions. Peterson, et al. write, “Our results illustrate the successes of human ingenuity, in particular, finding good functional forms for the EU and PT models [EU, Expected Utility; PT, Prospect Theory]. However, they also illustrate that this ingenuity can be supplemented by an automated search over models given enough data, and that as the class of models becomes less restrictive and the dataset becomes larger, this automated approach begins to substantially outperform the best of decision-making developed by human researchers.”
Is machine learning smarter than the human variety? Not necessarily, but it can keep more balls in the air. ds
© Dennis Simanaitis, SimanaitisSays.com, 2021