A simple tool to start using AI for decision making

Original author: Ajay Agrawal, Joshua Gans, Avi Goldfarb
  • Transfer
Hello! We’ll start the month with a rather light but useful material, the publication of which is timed to coincide with the start of the Big Data for Managers course , which starts in mid-April. So, let's begin.

There are a huge number of authoritative opinions on the impact of artificial intelligence (AI) on the business of the near future. But on the topic of how exactly companies can start using it, much less is said. Our study and book begin with an analysis of AI into its simplest components. We suggest a way to take this first step.

Let's start with a simple idea: the latest AI developments are aimed at reducing the cost of prediction. AI improves forecasts, makes them faster and cheaper. It has become much easier to predict not only the future (What weather will be next week?) But also the present (How is this Spanish site translated into English?). Forecasting is the use of available information to obtain information that you do not have. If you have information (data) that needs to be filtered, compressed and sorted to get ideas that make decisions easier, forecasting will help. And now, cars can help.

Improved predictions help make decisions in the face of uncertainty, which is a common situation for business. But how to consider introducing a prediction machine into the decision-making process?

We taught this topic to graduates of the Rotman School of Management's MBA at the University of Toronto and talked about a simple decision-making tool: AI Canvas. Each element of the canvas contains one of the requirements for making decisions using a computer, starting with prediction.

AI canvas

Use it to understand how AI will help you make business decisions.

What you need to know to make a decision?

How are various results and errors evaluated?

What are you trying to do?

What metrics are used to measure success?

What data is needed to run the predictive algorithm?

What data is needed to train a predictive algorithm?

How can I use the results to improve the algorithm?
To explain the work of AI canvas, we use an example invented by one of the workshops on AI strategy Craig Campbell, CEO of Peloton Innovations, an organization that implements AI in the security industry. (This is a real-world example, based on a product called RSPNDR.ai that Peloton sells.)

More than 97% of home alarm cases turn out to be false. That is, their cause is not an attacker. The security company needs to make some decision: whether to call the police or security? Call the homeowner? Ignore? If the company decides to act, in more than 90 cases out of 100, it will be in vain. However, taking measures in response to an alarm means that if there really is a danger, then the security company will not leave it unattended.

How to understand if a predictive machine will help you? AI Canvas is a simple tool to organize the necessary information into seven categories to obtain the necessary solution. Let's look at an example of a security alarm.

AI Canvas: An Example Using AI to Improve Home Security

Prediction Predict whether an alarm went off on an unknown person or something else (i.e. true or false).

Compare the cost of a response to a false response with the cost of inaction in the case of a true response.

react or not in case of a signal.

Whether the correct decision was made when the alarm was triggered.

Data of motion sensors, heat, cameras for every moment during the alarm. This data will be controlled by AI.

Sensory data for a certain period of time and the corresponding data of the results of the operation (a real attacker or a false alarm); this data is used to train AI before launching it.

Sensor data and the corresponding response results (confirmed by an attacker or confirmed false response); this data is used to update the model during AI operation.
First, we’ll clarify what needs to be predicted. In the case of an alarm, you need to find out whether it is caused by an unknown person or not (false alarm or not). A predictive machine can potentially report this - after all, an alarm with a simple motion sensor is to some extent a predictive machine. Machine learning allows you to use a wider range of sensor data to determine what exactly you want to predict: whether the movement was caused by an unknown person. With the right sensors, for example, a camera that recognizes faces - people and pets, or a door lock that recognizes when someone is near the door, modern AI technologies provide more detailed predictions.

The prediction is no longer in “movement = anxiety”, but, for example, “movement + unfamiliar face = anxiety”. More complex predictions reduce the number of false positives, which simplifies the decision to send the guard for verification, instead of calling the owner.

The prediction cannot be 100% accurate. Therefore, to determine the size of the investment in improving predictions, you need to know the cost of a false positive in comparison with the cost of ignoring the present. It depends on the situation and requires a human assessment. How much does a call back to confirm the situation? How much does it cost to send a guard in response to an alarm? How much does a quick reaction cost? What is the cost of inaction if the attacker is really in the house? There are many factors to consider; Determining their relative value requires an assessment.

Such an assessment can change the essence of your forecasting machine. In the case of alarms, cameras around the house are one of the best options for determining the presence of an unknown intruder. But many people may find this uncomfortable.

Some will prefer confidentiality over false alarms. Evaluation sometimes requires the determination of relative values ​​and factors that are difficult to calculate, and therefore to compare. The cost of a false positive is easy to measure, the price of privacy is not.

Then, determine the action that depends on the generated forecast. This may be a simple “react / not react” solution, or something more nuanced. Possible options include not only the reaction of someone, but also the instant inclusion of remote monitoring of who is at home, or some way of contact with the owner of the house.

Action leads to a result. For example, a security company reacted and sent a security guard to check (action), which detected the intruder (result). In other words, looking back, we can see whether the right decisions were made at all stages. This knowledge is useful for assessing the need to improve predictions over time. If you do not know what result you want to receive, improvements will be difficult, if not impossible.

Part of the canvas — prediction, evaluation, action, and outcome — describes important aspects of the decision. The other part is three final considerations. All are data related. To generate a useful prediction, you need to know what happens at the time the decision is made - in our case, when the alarm goes off. In the above example, this includes motion sensor data and visual camera data collected in real time. This is the most basic input.

But to develop a prediction machine, first of all you need to train a machine learning model. The training data consists of sensor data for a certain period of time with the corresponding results for calibrating the algorithms underlying the forecasting machine. In this case, imagine a huge table, where each row is the time of the alarm, whether the attacker really was and some other data, for example, time and place. The richer and more varied the training data, the better your predictions will be. If there is no data, you will have to start a mediocre prediction machine and wait for its improvement over time.

Improvements will come from feedback. This is the data that you collect during machine operation in real situations. Feedback data is often generated in a richer environment than training. In our example, you can find the relationship of the result with the data received by the sensors through the windows, which affects how the movement is recognized, and how cameras capture faces - which is perhaps more realistic than the data used for training. So you can further improve the accuracy of predictions thanks to continued training on feedback data. Sometimes such data will be uploaded to a particular house. And in other cases, may extend to several.

Explaining these seven factors for each important decision of your organization will help determine whether AI can reduce costs or improve productivity. Here we discussed a solution related to a specific situation. To get started with AI, your task is to identify key decisions in your organization in which the outcome depends on uncertainties. Filling AI Canvas will not be able to say whether you need your own AI or you can buy a ready-made one from the supplier, but will be able to explain what contribution the AI ​​will make (prediction), how it will interact with people (assessment), how it will affect decisions (action), how success (result) will be evaluated, and what types of data are needed for training, operation and improvement of AI.

The potential is huge. For example, an alarm triggers a prediction to a remote agent. One reason for this approach is the sheer number of false positives. But think about it - if the predictive machines become so smart that there will be no more false positives, will the reaction and sending the guard be the right decision? One can only imagine alternative solutions, for example, a system for capturing an attacker in place (as in cartoons!), Which could exist with more accurate and high-quality predictions. In general, improved predictions create more opportunities for new approaches to security, or even for potential predicting the intentions of an attacker before they penetrate.

If you find the material useful, we will be grateful for your pluses. And for a more detailed acquaintance with the course program, right now you can sign up for a free open webinar , which will be held by our teacher Artyom Prosvetov on April 3 .

Also popular now: