The world’s leading publication for data science, AI, and ML professionals.

The Future of Decision-Making in AI

How not to end up in a Skynet, The Architect, Ultron or HAL9000 situation

Photo by Yuyeung Lau and Photo by Xu Haiwei on Unsplash
Photo by Yuyeung Lau and Photo by Xu Haiwei on Unsplash

Generally, I am not a fan of the terminology Artificial intelligence (AI). It is too broad, and non-technical minded people (who took the blue pill) imagine AI as a singular entity that makes decisions independently. Additionally, because AI is a popular term, I have seen examples where companies advertise themselves using AI when they are actually "just" using linear regression. Throughout the last 80 years, the term has gotten a bad rap in pop culture because of all the doomsday science-fiction stories and movies. Countless times we have seen science-fiction turning into science-faction, and with the advent of technology like the text generator GPT-3 by OpenAI, it sure looks like we are on track. So, will this also happen here?

Nope.

Well, at least not for now. The autonomous capabilities of self-driving cars are characterised by seven different levels of automation (or intelligence). Artificial intelligence is classified into three main classes: narrow, general, and super artificial intelligence. At the moment, the highest autonomous car level available for the public is level 3 (max level 5), where the driver in specific situations can disengage from the act of driving. With regards to AI, we are only swirling around in the narrow artificial intelligence class – despite it sometimes seeming like higher-level AI.

Photo by Damiano Lingauri on Unsplash
Photo by Damiano Lingauri on Unsplash

Decision-making in AI is accomplished through three mediums: Data, Modelling and/or Decision-makers. The holy grail is to have all three incorporated in the decision process, but sometimes you lack data, and other times the process is so complicated (or fast) that decision-makers are best left out (e.g. why phone operators is a thing of the past). The main thing is to understand when.

This story will focus on how one properly integrates the decision-maker, as there are plenty of great stories, papers, books, etc., written on how to do proper modelling and integrating data science techniques to support decisions.

How is the decision-maker affecting the AI-made decision?

Many problems are too complex for us to recognise a solution immediately. Simultaneously, we are lazy by nature, so we want to find a better (and easier) method to obtain a solution – enter the "AI" method.

Preferences of the decision-maker can be split into hard and soft preferences. Hard preferences are generally referring to those opinions that the decision or solution must obey. In modelling terminology, these are sometimes referred to as constraints—E.g. when minimising factory costs while still enforcing a certain production level. Note, hard constraints and hard preferences are NOT the same, as some physical limits (work hours available, money available) cannot be surpassed, but they are often treated the same in modelling. The soft preferences are the "would like to" kind of opinions, which often have conflicting nature, are a bit ominous, and the reason behind most discussions on decision-maker, operator, management and even board level of AI companies. Paradoxically, the difference between hard and soft preferences is also a decision that humans enforce – I would argue that optimally we shouldn’t distinguish between these, and all preferences should be considered soft preferences.

But how are these soft preferences enforced? Here, we distinguish between when the preferences are incorporated in the decision process: a priori, a posteriori, or interactive. The choice of method coincides significantly with the complexity of the problem, the amount of computing power available, and the time requirements for when to yield the decision. Note, for some of these problems, obtaining a decision is also referring to the notion of obtaining a solution.

The common goal is to identify the solution that best represents the preferences of the decision-maker. At least, if the decision-maker knows their preferences – often they are very uncertain, fuzzy, a large group of decision-makers have to agree, or maybe they are straight up incompetent to define them (Que data and modelling methods).

Photo by Greg Rakozy on Unsplash
Photo by Greg Rakozy on Unsplash

A posteriori approach

For multi-objective problems where preferences are very specifically defined, a graph with non-dominated solutions are often created highlighting a Pareto front. Hereafter the decision-maker evaluates which of the solutions at the Pareto front that best reflects their preferences.

The method at large consists of computing a high number of solutions or simulating decisions in order to then evaluate the outcome and compare them to the preference structure of the decision-maker to then finally chose one. The process is rather extensive, as a solution has to be computed for a high number of combinations between different objective settings. And for many-objective problems, almost any solution computed will be a non-dominated solution (with respect to a single objective, it is the optimal solution), so the method quickly runs into dimensionality problems. But for smaller problems, it is nevertheless doable and super useful.

Example of Pareto front. Illustration by author.
Example of Pareto front. Illustration by author.

In these methods, the decision-maker has what is referred to as "posterior articulation of preferences."

A priori approach

This approach tries to combine objectives into a single objective before starting to find a solution. The problem is that one doesn’t know the effect of the problem on the combinations, how the uncertainty affects the outcome, and lastly, how to combine the different preferences.

Combining the preferences, one has to discuss whether criteria are independent of each other or whether there is a positive or negative interaction. For example, the two criteria "activity at a bus stop" (probability of a bus arriving within x minutes) and "the walking distance to the said bus stop" are dependent when deciding which bus stop to walk to. A higher activity level would make distance less of an issue.

Additionally, the preferences of certain criteria aren’t necessarily constant and perhaps not even linear. In the criteria "walking distance to a bus stop", a difference between two alternatives less than 10 meters doesn’t affect one’s choice – we are indifferent to such a difference. Meanwhile, a difference of more than 4 km would veto a bus stop, forcing one not to select that bus stop no matter how great the other criteria are.

Politics is a good example of how multiple decision-makers are trying to get together and through a priori approach tries to enforce a solution without knowing the actual outcome. Photo by Frederic Köberl on Unsplash
Politics is a good example of how multiple decision-makers are trying to get together and through a priori approach tries to enforce a solution without knowing the actual outcome. Photo by Frederic Köberl on Unsplash

That is, a priori approaches cover methods that estimate a value function, introduces "outranking relations", the analytical hierarchy process, or even tries to identify decision rules. In these methods, the decision-maker has what is called "prior articulation of preferences."

Perhaps, the most well-known value function method is goal programming; This method incorporates penalties for not reaching a prespecified goal and seeks to minimise the penalty. Thus, transforming the multi-objective framework into a single-objective one.

The point is; there is a lot to consider when utilising a priori approaches, but they are fast and applicable for large complex problems.

So why aren’t we gonna end up replicating scenes from The Matrix, I, Robot, Age of Ultron, Terminator, or 2001: A Space Odyssey?

The technology isn’t far enough (yet), and extensive research is being done on how to actively avoid this and properly incorporate decision-makers, modelling and properly train models on data that aren’t biased.

The science fiction writer Isaac Azimov proposed three rules that robots and AIs must follow. Namely:

  • First Law: A robot may not injure a human being or, through inaction, allow a human being to come to harm.
  • Second Law: A robot must obey the orders given by human beings except where such orders would conflict with the First Law.
  • Third Law: A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

But if they aren’t perfect, a super-AI should be able to find loopholes in them. The added zeroth law expresses one such fixed loophole. Consequently, we must promote research and technology that improves the collaboration and understanding of Decision Making in AI.


I hope you enjoyed this very general story on decision making and AI. Note, I didn’t go into details with characterising problems and which methods apply to which problems, thus also having omitted an essential discussion on evaluation problems vs design problems. If you are interested in other fields of research that investigates this, a good example is Operations Research, which I have written a small story on:

Why Operations Research is awesome – An introduction


Related Articles