Decision Making in Uncertainty

Applying artificial intelligence (AI) to personal and business decision-making problems depends on how AI can perceive and handle uncertainty and risk. There are 4 major types of uncertainties in decision-making problems.

• Data Uncertainty: We always had and will have uncertainty in our data. Noise in data is the most well-known type of data uncertainty. For example, sensors are not perfect, and they always record some degree of noise. It is important to know that uncertainty in data is not limited to noise. Uncertainty about the data source and what data is representing are also two other examples of data uncertainty. Imagine you have a perfect sensor, but you are unsure about your sensor's location (for example, in a gas pipeline or underground). In this case, you are facing data source uncertainty. In another case, suppose you are collecting data for a political campaign. You gather information from thousands of individuals but you are not sure how much your collected data represents the target community. In this case, you are dealing with uncertainty in data representation. Data uncertainty is crucial in decision-making applications that work with different sources of data, especially sensors. The field of autonomous cars is full of examples regarding the effects of data uncertainty on decision-making.
• Prediction Uncertainty: Using the available data, you generate some cases and try to predict the outcome of each case. For example, you want to buy a house. You collect the local real estate market data and your financial data. Based on the data, you develop some cases, such as buying a small house close to the downtown or buying a bigger house in a suburb area. After making cases, it is time to predict what happens in each case. If you buy a house in a suburb, you have more room, access to nature, more friendly community, but your house appreciation will probably not be too much. In the other case, you might have a smaller house, access to more restaurants and nightlife events, and your house will appreciate significantly in the current market. But how do you know those things? How confident are you that a house in the downtown area will appreciate more than a house in a suburb? How sure are you if your suburb neighborhood is more friendly than your downtown option? If you look closely, there are many uncertainties in your predictions, referred to as “Prediction Uncertainty”. Data uncertainty could be a source of prediction uncertainty, but it is not always the case. Even with the perfect data still, an unexpected event can change the course of events, and your prediction could go wrong. Imagine those who bought houses in downtown areas to be closer to their offices, and COVID-19 made many of those jobs remotely. Now, being closer to a closed office does not mean less commute time as they predicted in their house-buying decision. In this case, the environment changed dramatically, in such a way that the most accurate collected data could not predict it. AI decision support systems in the financial sector deal with this type of uncertainty, prediction uncertainty, in their decision-making problems every day.
• Judgment Uncertainty: You made some predictions based on the available data, and now is time to make a judgment. You must compare them against each other and rank them based on your utility function. Your utility function is a mix of financial, emotional, and many other factors. Going back to our buying a house example, your utility function is a mix of money that you pay, anticipated appreciation, comfort of living, enjoyment of being in a neighborhood, and many other factors. Your judgment should help you to rank different cases based on your utility function. But how can you value the comfort of living in a big house compared to the price that you pay for a house? How can you value a good neighborhood versus anticipated home value appreciation? Should you give more weight to 10% better neighbors or 3% more home appreciation? When humans or AI try to make a judgment between different cases, there is an inherent uncertainty due to this mixed bag of factors. In some problems, it is very hard to give a numerical value to some emotional factors and then compare them. Judgment uncertainty is more common in decision-making problems that are dealing with both financial and emotional factors. For example, product recommendation systems must solve judgment uncertainty before deciding on recommending a product to a consumer. Such a system should judge to recommend what product to whom to satisfy both his/her emotional and financial needs.
• Action Uncertainty: We tried to collect perfect data, predict outcomes accurately, and make the optimum judgment based on the best-perceived utility function. It is time to make a final decision and take an action. But wait! is there any uncertainty associated with taking action? Let’s see what action uncertainty means first. Action uncertainty means we decide to take a specific action, but in reality, we do a different action. It looks strange that how can someone decide to take a specific action, but the actual action will be different? I some cases, like buying a house, it is very unusual that you decide to buy a house and finally buy another house by mistake! But, there are many examples of action uncertainty in the real world. For example, you are about to hit an object with your car. You decide to stop. You must hit the brake, but unfortunately, you hit the gas pedal. In this case, the action was not intended by the decision, but it happened due to a mistake. The action uncertainty is not always due to a mistake. In the driving accident example, you might hit the brake pedal correctly, but because of bad tires, the car slides and hit the object. There is always a chance of action uncertainty that you must take into account. Action uncertainty is one of the important uncertainties in the fields of robotics.