You’ve heard of it, and it is powerful – some even perceive it as threatening. Artificial Intelligence or AI is a term often used, yet little understood.
The engineering truth behind the term »Artificial Intelligence« is a set of tools that may do jobs better or handle[ds_preview] tasks we could not handle at all in the past. Key techniques of AI are machine learning, knowledge-based systems, natural language and gesture processing, and generally robotics. The following gives a short extract from a Hiper 2020 paper (www.hiper-conf.info).
Machine Learning
Machine learning and data mining are closely related to computational statistics. In essence, we have glorified statistics here with new, catchy labels attached to it. Most semi-empirical methods in maritime applications, such as ship design, have approximated data points by a constant or at best very simple function. Wouldn’t it be nice to have some mathematical way of mimicking the curve we would instinctively draw through such data sets, ignoring implausible outliers and following the trends our eye sees; something flexible yet smooth and free of inappropriate oscillations?
For the naval architect, this is old hat. We have approximated arbitrary point sets for centuries, first using flexible thin beams (splines), and later using aptly named spline curves, which do not oscillate and form smooth curves and surfaces. The machine learning community prefers other functions, such as sigmoid functions. Combining many of these, we have similar basic qualities of flexible approximation and avoid oscillations.
The fitting process is performed most often by Artificial neural networks (ANNs). These represent functions by »layers«, and coefficients by »nodes«. More nodes and more layers offer more flexibility in fitting. And if you have more layers, you use the term »Deep Learning«. Take away the glamorous terms and you end up with a plain, but useful statistical tool. There are countless applications for ANNs, as pattern matching or trending is needed in many fields: fingerprint matching, facial recognition, speech recognition, automatic reading of licence plates, playing chess or Go. Useful maritime applications include empirical models for ship design, performance monitoring or pattern recognition, e.g. automatic ship identification or spotting deficiencies in drone inspections.
Not magic
But ANNs are not magic. They can’t predict the unpredictable, such as random events (including the effect of random events on ships). And they aren’t good at predicting highly nonlinear events, as we have random errors in the initial conditions which lead to a large scatter in the results, e.g. for crash-stop manoeuvres. They are also data greedy and if you have only a few data sets and not thousands or millions, as is typical in many maritime applications or for a rare event, we should use natural intelligence, e.g. reducing the number of free variables using physical insight.
Knowledge-based systems
Knowledge-based Systems (KBS) or expert systems reason within narrow knowledge domains. Most KBS use IF-THEN rules to represent knowledge. Rules may be taken from regulations or from experience, e.g. interviewing experts. KBS become powerful when you have many, many rules which become too complex for humans to handle, or when quick response is needed.
KBS incorporating uncertainty or probability are labeled as ‘Bayesian networks’. Case-based reasoning (CBR) systems are special KBS using an approach similar to that of lawyers or doctors: Find related cases and study them to derive the best strategy for the current case. (Intelligent) agents are swarms of simple expert systems working (communicating) together.
The challenge
The challenge lies in getting all relevant rules together. This task can get (prohibitively) daunting if expert knowledge acquisition takes too long or costs too much money, if rules change rapidly or if there are no agreed rules, such as in creative fields like (aesthetic) ship design. If the KBS does not contain enough rules or cases, it is often useless for practical purposes. Most maritime KBS applications are found for ship operation. This is not surprising; ship design is a creative and complex process in a rapidly changing economic and regulatory environment.
In contrast, collision avoidance follows rules (COLREGs) that have hardly changed in the last 100 years and rules for emergency response are conveniently documented in emergency response plans and handbooks. DNV GL works on combining machine vision and case-based reasoning for next-generation plan approval. The idea is to identify similar plans to the one submitted for approval, retrieve the corresponding cases, and derive recommendations based on these.
Speech Recognition
Speech recognition and its shortcomings are widely known. Just google »speech recognition gone wrong«. But the advantages of hands-free operation and handling simple user interaction by voice are too compelling to ignore. Speech recognition is generally built upon machine learning, both for individual words and commonly used word combinations.
Applications of speech recognitions range widely, including control of secondary equipment in airplanes and cars, mobile email, etc. These are ubiquitous in all our lives. As a maritime application, the voice-operated Super Bridge-X system of Mitsubishi allows in principle »no-touch« operation of the ship.
Useful or dumb?
Like other computer tools, AI can be very useful, but also incredibly dumb and stubborn. We should use AI more; we would be dumb not to exploit its capabilities. But we should also be aware of its limitations and use the tools wisely.
Volker Bertram
