Using the right tool for the right job makes the job a lot easier. If we’re lucky, that’s what we learned as children. The key to productivity and effectiveness is not letting the tools you use get in the way of accomplishing your task. Not to say that one tool is intrinsically better than the other; just that one tool is more appropriate for one job than another.
The same can be said for machine learning algorithms. It’s not that one approach is better than the other per se. However, if you use the right algorithms, or combination of algorithms, you will be a lot further along to solving your problem.
Here’s an example. Say that you want to predict the possibility of a type of industrial machinery failing (and as a reader of this blog, there’s a good chance this is true.) You’d want to know what might cause a failure, and what states or conditions might the asset pass through on its way to failure.
Hidden Markov models would be a good algorithm to infer and track back the highly dimensional hidden states or conditions the asset might go through on the way to its unfortunate demise. Why? Because despite all the known history of the asset in use across the industry, not every path to failure may be known. Most current approaches to predicting the possibility of asset failure rely on overly simple assumptions just to make solving the problem tractable. This typically results in a single, stationary distribution for analysis, albeit a well-formed one. However, the majority of industrial machines are hierarchical in nature. Their failure process is described by a blend of many simple distributions that form a more-complex and highly-realistic probability distribution. An unsupervised learning algorithm like Hidden Markov models will excel at discovering all the hidden paths to the asset’s failure.
Seems basic, yet many companies look for asset maintenance solutions that don’t use the right tools for the job. These solutions base their predictions on a statistical average across an asset class. These only seek to predict known causes of failure. Depending on your asset (and the complexity of your maintenance operations, the cost to repair or replace, the ready availability of spare parts or replacement assets, etc.), a statistical average approach might be good enough. But if your assets are core to meeting your production or service commitments, it may be worth researching other options.
You can start that research by downloading our white paper (registration is required).