Artificial Intelligence (AI) is the latest ‘gadget’ that companies want to add to every project. However, it’s not the silver bullet that many folks hope it to be. It has its strengths and limitations. To overcome AI’s limitations, Augmented Intelligence optimizes the strengths of Artificial Intelligence, Machine Learning (ML), and human capabilities.
The fundamental limitation of AI is not the programming, but the design of the AI algorithms. Artificial Intelligence is a rules-based solution, and a traditional rules-based AI requires complete and accurate rules. It’s up to the designers to accurately define the right rules. Historically, humans are not very good at predicting every eventuality or possibility, and thus do a poor job of specifying the rules.
And then there’s the issue of the data used to train and operate an AI system.
Humans routinely arrive at successful solutions based on ambiguous, inaccurate, and incomplete data. Computer and AI systems require accurate data to derive an accurate solution. Unfortunately, inaccurate data is a common occurrence, and it’s not always possible to identify the inaccuracies. For instance, there may be inherent biases in the data that will alter the results. These biases may be created by virtue of how the data was collected or due to an unrepresentative data set.
Accurate data is critical, especially if the tool includes a Machine Learning component. A learning engine is only as good as the data it learns from. A good AI solution should include a learning engine to constantly evolve its rules based on actual usage.
Creating a training data set requires a lot of human processing, which may account for much of the unintended biases in the data set. A human decides what data to use to train a learning engine. Without knowledge of how the training algorithm works, they may exclude data that would otherwise be useful and include data that confuses the learning engine.
Besides data accuracy issues, AI systems are not designed to understand inferences as well as humans. An AI engine might be good at taking ice cream orders, such as a chocolate ice cream cone with candy chunks on it, but the same engine would not know how to handle a request for a cone, “Just like that one only with sprinkles.”
AI systems are designed and trained to perform specific functions with specific data sets. They are not yet sophisticated enough to generalize across all inputs to learn everything. Humans are uniquely capable of transferring knowledge across domains to operate effectively in a novel environment based on what we have learned in other, non-related environments.
Humans have limitations, as well. Humans just don’t have memory processing capabilities as good as computers. We are not able to process large amounts of data, we cannot keep track of many things at one time, and we cannot recall things with perfect accuracy. These human limitations are AI strengths.
So, one of the big questions is how to leverage the strengths of humans to overcome the weaknesses of AI, and vice-versa. Augmented Intelligence is a model that emphasizes the assistive role AI can have in enhancing human cognition rather than trying to mimic or replace it. A good example of the use of Augmented Intelligence is in the Automated Readiness Forecasting (ARF) tool designed by the AF CyberWorx team.
ARF accepts mission parameters, such as an exercise several months away, and schedules qualification events to ensure that appropriate personnel and resources are available for the exercise. It also keeps track of all necessary equipment maintenance as well as personnel training and medical activities needed prior to the exercise. It suggests schedule changes which will ensure the required resources are ready in time. Affected personnel only have to approve or acknowledge any adjustments. The tool then tracks progress, highlighting any deviations to the plan and suggesting corrections along the way to stay on track. Machine Learning uses the repeated iterations and changes to evolve the scheduling algorithm, fine-tuning its capabilities and reducing reliance on human intervention.
The readiness tool assumes the mundane tasks humans would otherwise have to perform to keep track of readiness data and immediately adjusts the plan to accommodate any changes to resource availability. Normally, such adjustments would require dozens of people working hours to accomplish. Humans are still required to approve changes, but are relieved of the time-consuming and error-prone tasks associated with adjusting schedules manually. This tracking and automation saves hundreds of resource hours, eliminates errors, and ensures a higher level of readiness.
This example illustrates an Augmented Intelligence and Machine Learning paradigm that could be applied to many Air Force projects rather than a typical Artificial Intelligence approach. The point is, AI is not a panacea for all problems. AI is successful when the problem domain is well understood, the data set is accurate, and the AI is focused on a specific task. Knowing when to design for an Augmented Intelligence verses an Artificial Intelligence – and the difference between the two – is crucial to success. Not every problem needs an AI solution.
*The postings on this blog reflect individual team member opinions and do not necessarily reflect official Air Force positions, strategies, or opinions.