Raising the bar in Artificial Intelligence and automation
 
 
Many manufacturing processes are now automated on production lines, but what happens when there’s a level of human planning, skill or judgement required?
Until now, many actions and processes that require higher-level thinking have been seen as too difficult to automate but WMG’s Centre for Applied Artificial Intelligence is undertaking research that could change that.
 
Dr Malcolm Strens, honorary fellow at WMG, explains the Artificial Intelligence (AI) technique that he thinks will transform the factory of the future:
 
 
A tool for transformation

AI has been around for a while but it has come to be associated with neural network technologies, where machines learn to perform tasks automatically, rather than being programmed with a set of rules.
It’s important for manufacturers to be aware of developments in AI because the UK is involved in a global race to apply these technologies and improve productivity.
In around 15 to 20-years, most of the simple, dull, dirty and dangerous industrial tasks will have been automated. As a result of breakthroughs in AI, it’s possible that much of the decision making and planning, currently undertaken by managers, will also be done by computers and machines.
 
The most promising branch of AI

At WMG, we are conducting research using reinforcement learning, one of the main branches of machine learning that will underpin these changes. Reinforcement learning is being heralded as the ‘true hope’ of AI because of its huge potential. We expect that it will dramatically increase productivity by enabling the automation of more complex systems, processes or tasks that are currently seen as being too difficult to tackle.

UK based AI company, Deepmind, now owned by Google, have led the way in demonstrating that reinforcement learning is a powerful technique. It could change the way a lot of businesses operate and may lead to the automation of a lot of things that we thought were going to be left to humans for quite a while. That includes tasks that require higher level thinking, judgement and planning.
 
Trial and error

Reinforcement learning techniques are applied to processes that require judgement and planning, where it’s necessary to go through a series of steps before it’s clear whether you have been successful. This branch of AI differs from other techniques because the system (or learner) in is not told what actions to take but learns to discover the actions that lead to the biggest reward. It’s a trial and error process that’s like the way a child explores their environment without supervision.

Creating a digital twin

We can’t learn by trial and error in a chemical plant or a car factory, so when we apply reinforcement learning in manufacturing, we start with a simulator or “digital twin” of the production process.The learning algorithm or ‘agent’ learns by trial and error and does not require a teacher, only a measure of success (the reward). Once an effective strategy has been learnt, it can be used in the real system.

We are open to meeting with companies that are interested in learning more about applying this technique.
  
 
 
Newsletter navigation: