An earlier blog post outlined the application of machine learning within the concepts of ‘Big Data’ and the ‘Internet of Things (IoT)’. How can a company get started with the exploration of the possibilities represented by machine learning?
Machine learning (ML) is a current business buzzword, located at the top of the ‘Peak of Inflated Expectations’ on Gartner’s Hype Cycle for Emerging Technologies for 2017.
Machine Learning split out from the wider concept of Artificial Intelligence (AI), when AI went in the direction of Neural Networks, while Machine Learning went more in the direction of pattern recognition and prediction from data. It is closely related to statistical analysis, such as principal component analysis and ‘Data Mining’. The analysis can be unsupervised (with no preconception of what sort of relationships are expected), or supervised, where the relationships are checked against expectations and ‘Reinforcement Learning’ applied to improve the predictive capabilities.
So, how to proceed?
Machine Learning is a collective term for many different types of algorithms, and there are many different suppliers of software. There are commercial systems, pay as you go, and freeware. Large computing companies, such as Microsoft and Google, have developed systems and there are community-supported freeware systems. Aspects affecting the choice of software system include data storage and data security, capability, ease of use, integration with control and information systems and cost, amongst others.
For a large process industry operations, the number of process streams and sensors can be large, and the amount of data further increased by storing values at short time intervals over many years. Hence for large companies the ability to store data on cloud servers is an advantage, as well as software capable of handling large data sets, with fast operation via access to remote computing resources. A popular example of a cloud-based ‘Software as a Service’ (SaaS) system is Microsoft Azure. This offers a gratis profile, with charges for just the features that you use. The prices are outlined on the webpage, but it can be difficult to estimate in advance which features are needed, and hence the total cost can be difficult to estimate in advance. Data security naturally has high focus, but apparently the US government can be allowed access to all data, even for foreign companies storing their data outside the USA. However the ease of integration with other Microsoft software, and the pay-per-use makes Azure a popular choice. For SMEs looking to provide Machine Learning capabilities to process industry the choice of a widely implemented and well-known platform saves resources learning or introducing different systems for each customer.
On the other hand, businesses wishing to explore the capabilities of machine learning without up-front spending on software could start with free and open-source software. This can require more involvement in terms of development work or programming, but the open-source software can also allow a greater degree of tailoring of the system to the customer. An example of an open-source platform for machine learning is Weka, the popularity of which is evidenced by a review of the textbook having received over 30 000 citations in scientific literature. The Weka platform is Java based, allowing tailoring of integration with other information systems. Instruction resources, with a hands-on approach are available via Youtube instructional videos.
A review of recent progress in the application of machine learning in process systems engineering highlights the advances made possible by reinforcement learning. IBMs Deep Blue chess computer managed to beat Garry Kasparov in 1997 using a brute force approach (searching to a depth of 6-20 moves), but the Google DeepMind computer victory over Go champion Lee Sedol in 2016 required a different approach, given the number of possible moves in Go being beyond the (current) computational limits (10360). The DeepMind approach combined supervised learning with reinforcement learning, and enabled the computer to engage in self-play, and further improve over time.
The success of the DeepMind/Deep Learning approach shows that pure statistical number-crunching (unsupervised learning) will not be as powerful as algorithms that incorporate supervised or reinforcement learning. Correlations between data can be due to real relations, or a result of coincidences. Human involvement to review the relationships can be needed. Within scientific methodology there is a fear of the erroneous inference of causation from correlation, as illustrated by this scary Halloween decorated pumpkin (source: facebook). Hence human input to process systems engineering will not be made redundant by machine learning, at least for a while….
Ross Wakelin
Northern Research Institute Narvik A.S.
ross@tek.norut.no, (47) 99 252 485
