Logical Thinking AI that interprets the result logically with reasonsFeatured Technologies
December 12, 2018
NEC successfully developed the Logical Thinking AI―what is it like and in what circumstances is it useful? We asked the central person on the research team to find out.
AI presents hypotheses and provides convincing reasoning for them
― How is the Logical Thinking AI different from current AI?
The Logical Thinking AI tries to leverage human wisdom well. We are working on the development of putting AI that can assist human decision making to practical use.
For example, think about a case when some kind of problem occurs at an industrial plant and needs a quick decision. Suppose that plant had introduced an AI that assists safe operation. Conventional AI can only present a conclusion along the lines of "Close valves A and B." The reason why we should do that is not explained. This type of AI is only producing hypotheses after learning from numerical data; which means that this suggestion may very well be seen as bizarre by the people following the standard modes of thought and reasoning on site. Under such conditions, we humans have a hard time believing the AI and acting on its proposals. If the hypothesis proves to be incorrect, there is a risk of major damage to the entire plant, and it is us humans who bears the burden of responsibility in the end.
In contrast, a reasoning AI not only learns from numerical data, but also learns human wisdom based on plant operation manuals and other procedures. This enables the provision of a convincing reason for the suggestion to "close valves A and B": By referencing pages so and so of the manual. The referencing of wisdom from operating manuals also ensures that valid hypotheses are presented without deviating from the rules of the enterprise or team.
Not to mention, the conventional machine learning is also performed solidly in parallel. For instance, plant operation also needs the know-how of experts that cannot be verbalized in a manual, such as how much efficiency change the turn of a valve brings. The accuracy of such parts can be sufficiently improved by machine learning using simulation.
Maybe you will have a better idea if I say that a reasoning AI is like a detective doing deductions. For example, if the detective just pointed out who the perpetrator is at the end of a mystery novel and that was the end of the story, no one would be satisfied. Our AI can present a convincing line of reasoning―along with the logics and evidence―for why someone can be considered a culprit.
Significantly reducing machine learning time by logical reasoning
― What technologies are embodied in Logical Thinking AI?
Actually, the idea for this technology finds its roots in a technology called the expert system, which was commonly researched worldwide in the 1980s. The expert system tries to reproduce experts' decision making on a computer, and it was the precursor for the technologies attempting to use human wisdom. To apply this system to large-scale problems, a vast amount of computation is required―an amount that is so large it needs a quantum computer. For this reason, it was not suited for large-scale plant operations, and therefore it was not adopted at the time.
Now, we aim to develop this technology to a practically applicable level while drawing on the essence of the expert system. We succeeded in creating new value by combining machine learning with the expert system idea.
As mentioned earlier, to reflect unverbalized expert know-how in the system, we need to improve the accuracy to a satisfactory level by means of machine learning using simulations. However, conventional machine learning takes into account all possibilities and repeats prodigious amounts of trial and error with a simulator, which required years of learning time. On the other hand, this technology uses human wisdom such as operating manuals to refine the simulator's trial and error to an optimal range. This led to a great reduction in learning time to a matter of days. Finally, practical application is becoming a reality.
To conclude, the Logical Thinking AI is a system that combines the utilization of human wisdom―the expert system―and data use by machine learning, which is seeing increased application nowadays, to solve complex, large-scale issues.
Using reasoning AI at large-scale plants and other sites where human decisions are prompted
― What kind of applications are you thinking about?
At this point of time, we have chemical plants and other large-scale plants in mind. We are currently working on experiments for stabilizing quality in view of external disturbances such as weather. We also anticipate that this technology will be useful for human resource cultivation. While it takes time for people to be experts, using this technology can formulate efficient operation plans by using a simulator to significantly shorten that period.
Nowadays, AI tends to attract attention particularly in the field of automation in pursuit of streamlining. Of course, that is also an important theme and a field that we conduct research in. Notwithstanding, however much AI widens its application to automation, there are going to be situations where humans must make the final decision. A certain degree of risks always exists in the real world. Those risks are in the realm where humans must bear the final burden of responsibility and make the decisions. We believe that AI truly finds its practical application in the real world when it can present solid logical reasoning for such situations.