01
Dec
Maritime meets AI – Part 1
in All
Comments
After what could be described as a slow start in the maritime sector, the ways in which AI is benefitting shipping are now growing steadily. Current applications include the reduction of costs (from the better management of fuel consumption to identifying the most efficient routes), improved cargo management (e.g. the monitoring of shipping containers to prevent spoilage), and enhanced safety (e.g. identifying ships in close proximity and the safe usage of busy shipping lanes).
In our own products, both existing and those in development, we are making use of AI wherever possible. The rationale behind CompassAir has always been to empower those using our software to work in the most efficient way possible, enabling them to devote more time to those tasks that add the most value. Moving on from the efficient handling of what can amount to thousands of emails a day, the natural next step is to have the software read those emails, and having understood what they contain, based on experience (machine learning) to establish and assist wherever possible with the next, best steps.
An example of this is our new AI product for SnP brokers, replacing our longstanding ShipLink . Having learned from literally hundreds of thousands of emails, the software is able to read emails and understand their content. When, for example, it spots a Purchase Enquiry it will scan a broker’s vessel database, find vessels that match the P/E criteria, check which are available for sale, and then make suggestions to the broker. Bear in mind all this takes a matter of seconds and is running 24/7: an email arrives and almost immediately the broker is notified by means of a message containing vessel recommendations, even outside of normal working hours.
In the next article we will take a closer look at other ways in which we utilise AI, as well as how AI is currently being used in shipping generally, and how this is likely to change in the future.
First however, in order to put the subject into context, we will provide a very brief insight into some AI basics, in particular some of the more commonly used terminology and techniques.
What is AI, is there an agreed definition?
It was shortly after WW2 when interest in building intelligent machines started to take off. One of the most widely recognised names from that era is that of Alan Turing, the English mathematician and computer scientist, who in 1950 posed the question “Can machines think?”. In his article “Computing Machinery and Intelligence” Turing considered how a machine might be defined as being intelligent – from which was derived the concept of the “Turing Test”.
The idea behind the Turing Test was that for a machine to be considered intelligent it should be able to deceive an observer into thinking it was human – this was envisaged with an observer interacting with a machine and a human, with both attempting to convince him or her that they were human. With regard to intelligent machines, there is still no one, universally agreed definition, and to this day academics still don’t even have one single definition of AI. Broadly speaking it can be described as follows:
AI is a combination of computer science and data in such a way that allows problems to be solved, making machines do things that normally require human intelligence
Parts of the AI whole
AI encompasses a number of fields or branches, including machine learning (and deep learning), neural networks, robotics, and natural language processing. From these, the terms machine learning and deep learning will probably be the most familiar to most people.
Machine learning is essentially technology that learns from data, it looks for patterns within data. The more data available, the more that is learned. The better a dataset is understood, then from this it follows the better will be the predictions that can be made. The goal here is to make intelligent predictions based on prior data – the machine learns on its own to find rules, rather than rules being hardcoded, and these rules are then used to make decisions in new situations.
Supervised machine learning is where the model is trained – it is told what the right decision is for a given set of situations. Once trained, the model can be fed more data that represents different situations and then make intelligent decisions based on what has been learned. Perhaps the most common example of its use is in identifying spam – the more emails the system sees, having been told which are spam and which are not, the better the system will be at spotting future spam. Another example of its use, this time relevant to shipping, would be in approximating the value of vessels – the computer is fed data on the characteristics of a ship, its size, age, where it was built, etc. along with a label, namely the price for which the vessel was sold. From a sufficient number of datasets the goal would then be to predict the likely value of a new, different vessel, based on its specific characteristics, namely its size, age, etc. Closer to home, we mentioned our own CompassAir products increasingly making use of AI, software that can read incoming emails and extract structured intelligence, such as vessel details, open positions and cargo orders, having been trained using supervised machine learning.
Unsupervised machine learning on the other hand is where the label is this time missing from the training data – the computer has to work out for itself what it should be. In this case the learning algorithms must discover for themselves patterns in the training data sets. The advantages here include a reduced human workload being generated in both assigning and verifying labels; the disadvantages include the fact that more training data is required, with an increase in computational requirements, and the process is slower. Also there is a risk of distortion arising from anomalies in the training data.
Reinforcement learning is another way in which machines learn. This time they are not trained but instead learn from experience when interacting with their environment. Instead of learning the relationship between inputs and outputs, this time the agent learns by means of receiving a reward or punishment for the action it takes – eventually it learns what are the best actions to take in a particular environment. The goal of reinforcement learning is for the machine to learn a series of actions that maximise a long term reward. A practical example is Deepmind’s AlphaGo – it learned to play the game of Go by playing against itself, and then used what it had learned to beat the some of the world’s top players.
Deep learning is a part of machine learning, this time where the behaviour of the human brain is simulated using neural networks (see below). These networks are then capable of learning unsupervised from unstructured or unlabelled data. It is deep learning that is driving the huge growth we are seeing take place in AI applications, accelerated by the availability of increasingly powerful computers with simultaneously decreasing costs.
Natural language processing, another part of AI, relates to the interaction of computers with human language – the modelling of human language combined with machine learning and deep learning allows computers to process text and sound, and from this understand their meaning. Examples of NLP applications include digital assistants (Cortana and Siri), dictation software, chatbots and the like.
Robotics concerns the design, development and use of intelligent robots that combine both engineering (electrical and mechanical) and computer science. The functions integrated into robots include motion control, sense of surroundings (computer vision, object detection, etc), as well as the processing of training data, detecting patterns so that appropriate actions can be taken.
Expert Systems were one of the earliest forms of AI, first created in the 1970s. They solve complex problems by reasoning based on bodies of knowledge. They consist of a user interface, where the query is passed to the next component, the inference engine, essentially the brain. This then refers to facts held in the knowledge base, selecting information that will help in solving the problem using rules. Whilst the advantages include fast and efficient solutions to problems without constant reference to human experts, the disadvantages include the inability to find creative solutions to extraordinary situations (such as a human expert might provide) and, of course, if errors are present in the knowledge base then incorrect decisions will be made.
Artificial neural networks
When building and using AI systems there are many techniques that can be applied, one of the most widely known examples of machine learning being the use of artificial neural networks.
Neural networks are based on what is currently known about the workings of the human brain. A typical neural network will consist of a few to millions (a deep neural network) of artificial neurons arranged in a series of layers (the more layers, the deeper the neural network). Each neuron is responsible for solving part of a problem. The input units receive information from which the network attempts to learn, the output units provide the response to the information learned. In between are the hidden layers, the “brain”. The units are connected and weighted according to the influence they have on the next unit.
When the network is being trained, data is fed into the input units, then essentially a series of calculations take place. The network learns by backpropagation: the actual output produced is compared to what had been expected and the difference is taken into account in order to modify the weights of the connections between the units, working backwards to the input units. Over time, the weights are adjusted until the difference between the actual and expected output is reduced to zero – the program adapts as it is trained. Complexity can be added to make the system more robust, and eventually it will be ready to deliver an output from a new set of inputs.
An necessarily brief overview, but hopefully one that conveys the general idea.
Weak and Strong AI
AI is said to be weak or narrow when a machine can perform a specific task better than a human – intelligent behaviour but only for specific tasks. Search engines are an example – whilst very powerful, they are limited in what they can do: they give you an answer to your specific question. Another example is a robot on a production line – repeating the same set of actions, but only those actions.
Strong or general AI is where a machine has the same intelligence level as humans or better, this time across a wide variety of tasks. Only weak or narrow AI exists at present. When general AI is achieved – some predict this could be in 25 to 30 years from now – machines will have consciousness and be able to make decisions – machines would act and think like human beings. Whilst machines will still have to learn in the same way as humans, that is from experience, the difference is that this process will be many times faster than the time it takes us to learn.
AI in the maritime sector
Maritime is notoriously conservative, however the use of AI is becoming more widespread. Applications are already being used to make business processes more efficient, to improve the planning of voyages, decreasing shipping times, and to generally reduce costs, for example expenditure on fuel and on vessel maintenance by highlighting likely failures before they occur.
