What is a Utility Function?

Economics + AI Brains

AIs are not being built to simply do whatever they want to do. There are limitations and it all starts with the rules that are put into place when the systems first turn on. Utility functions essentially assign weights or ideal utility values to certain actions that the AI system can take. If you’re interested in an example of the typical math of a utility function, take a look at the link from Charles University’s, Roman Barták below, which explains it in detail.

All in all, utility functions seem to incentivize an AI to act in a way that caters to it reaching the desired outcomes of its actions as well as benefiting the environment, around it.

The question that remains is: how does this all play out in practice?

The Intersection that produces Utility

The usage of the utility function in AI systems seems to occur in a connection between math, the theory of reason, and probability theory, among other factors. Precisely, it appears that none of the function works without the above intersection of the three principal schools of thought. One can, in fact, begin by thinking of a utility function through an economic lens. In this case, it can be understood as the idea that every individual has goals, preferences, and reasons for what he or she does.

In the case of AI systems, all of this relates to their ideal outputs. When AIs are created, the teams that help them to grow and learn can train them with training data. This training practice is usually called “supervised learning.” While going through the phases of supervised learning, the team can effectively teach the system what values are ideal and what values are not.

For example, if a system is supposed to translate a text from one language to another, the AI team can input data that shows the system the percentage of error that it deal for success. In addition to this, this input data can also illustrate acceptable shortcuts that the system can take in the translation process. It also appears that utility functions can be used to input the values that the system should conclude with after each translation, in numerical form, in order to directly teach it how to perform at an optimal level.

Similar to this, utility functions can help an AI system assign numerical values to what is a desirable decision and what is not. If it achieves a desirable number, it will continue to do the action and if it achieves an undesirable number, it will not.

Even with these explanations, discovering the true place of utility functions in Artificial Intelligence can be difficult.

Von Neumann, Utility Functions and a Future Look

When you’re able to understand a bit of the work that John Von Neumann did for Artificial Intelligence, then it becomes a bit easier to understand how utility functions factor into space. Von Neumann was born in Hungary and moved to the United States to work at Princeton in 1933, after being offered a lifetime position there due to his fame as a mathematician and a researcher. John Von Neumann is not only known in the AI industry for his work with utility functions but also for what are called “Von-Neumann” machines.

Essentially, the theory of Von-Neumann machines introduced the idea of self-replicating robots, which has now translated to Nanobots that could one day replicate to a point at which they consume all Earthly matter. Despite this, Von-Neumann’s work with utility functions has also taken a significantly less drastic turn in the present time.

Overall, the growth of this theory began when Von-Neumann got together with Oskar Morgenstern to pen a work of research in economics and game theory called, “The Theory of Games and Economic Behavior.” Inside of this work lies the basis for the current usage of utility functions in Artificial Intelligence systems. This is, namely, the Von-Neumann-Morgenstern utility function. For the purpose of clarity, at this point, we’ll skip ahead to the connection between the utility function’s conclusions and how AI systems are currently structured.

Encyclopedia Britannica summarizes the idea of a utility function by stating that when a customer needs to make a decision that involves several options that must be left up to chance, then he or she will choose the option that he or she thinks will maximize his or her satisfaction related to the decision. In the case of AI systems, this has been extrapolated as if an AI system faces a decision involving a set of data that it feels reasonably uncertain about, then it will choose the ideal input values that best satisfy its utility functions.

With that, we have it.

Utility functions, overall, help AI systems make decisions in the most numerically efficient way that they can. In future pieces, we’ll delve further into this including into how this is being put into action by specific companies.

During this time, we’ll attempt to provide specific examples to show you how the ups and downs of utility functions play out.