L. L. Thurstone, an American psychologist, introduced an idea in 1927 that was revolutionary for its time — laid out in his paper, “A law of comparative judgment.” Essentially, Thurstone said that when people are presented with several choices, they invariably go for the one they value the most, even if they can’t quantify exactly why.
Thurstone was a pioneer in the field of psychometrics, which is centered around the idea that even mental processes that can’t be seen or touched can still be measured and understood scientifically. This set the stage for the later development of random utility models (RUMs), mathematical frameworks that help illustrate and predict human preferences.
RUMs can be used to evaluate the benefit or “utility” derived from a specific choice. Let’s say you’re picking which book to read next from a stack at the library, a RUM might help predict your choice. These models have an element of randomness to them, which takes into account the fact that individual preferences can be quite fluid and can change over time. As Gabriele Farina, an MIT assistant professor, explains, one might prefer coffee in the morning but, by evening, opt for a cup of tea instead.
These models have found applications in a variety of domains ranging from government to industry, aiding in predicting behaviors in scenarios like optimal transportation choices during construction or the best ways to allocate funds for public benefit. However, despite their widespread use, there’s always room for improvement in these models.
Recently, a revealing paper was presented at the International Conference on Learning Representations in Rio de Janeiro. This paper, authored by a team that included Yeshwanth Cherapanamjeri, Gabriele Farina, Constantinos Daskalakis, and Sobhan Mohammadpour, showed that traditional methods of making estimations, which rely largely on comparing two choices at a time, had definite deficiencies. They found that such pairwise comparisons make it hard to see the correlations that often exist between multiple choices. Recognizing these correlations, they argue, is key to making accurate preference estimations.
One significant finding of their research is that correlations can be discerned better when people rank three alternatives instead of two. This results in a more comprehensive understanding of preferences and gives a clearer overall picture.
The focus of this group from MIT is on the computational side of RUMs with a keen interest in developing algorithms that extract preference data efficiently. Emma Frejinger, a computer scientist at the University of Montreal, highlights this paper for its mathematical proof of why traditional data collection methods fall short and also for demonstrating the potential of using the best-of-three choices for accurate model training.
The research team believes that building utility models will continue to be an important focus in this field. These models have been critical to internet-based businesses starting from the late 1990s and will continue to play an important role in fine-tuning AI models. Their utility extends to language models (LLMs), where ranked preferences during training are employed to improve the model’s performance.
In a world packed with more choices than ever, it’s unrealistic to expect people to shape all their preferences for every scenario. Instead, models must predict preferences and continually evolve to make better predictions. To dig into further details, check out the original article here.
If you’re looking to leverage AI automation for your business, explore the myriad possibilities with implementi.ai.
This website uses cookies.