IBM Research

Data

The world’s newest natural resource

Machine learning, AI and cognitive computing have the potential to revolutionise the way in which computing affects our everyday lives. At a time when we are producing more data than ever before, it becomes essential to develop new and improved algorithms to shorten the time to insight from these exponentially growing data sources.

Much of the data produced outside of the social media footprint is produced by industry, and our efforts to develop capabilities that can allow industry to reap the benefits of advanced machine learning, AI and cognitive computing methods is well aligned with UK Digital Strategy and the recent report on Growing the AI industry in the UK.

“Our mission statement is simple: To deliver state-of-the-art research solutions through consumable, reusable technology that has real-world value to industry.”

—Edward Pyzer-Knapp
Research Lead, Machine Learning and AI

Intelligent simulations

AI in high-performance computing

Computational scientists have long relied on Moore’s Law (bigger, faster, cheaper) to achieve their complex computational workflows on ever more powerful machines. Recently, however, this bastion of the digital world has started to show some cracks. The required speedups are no longer guaranteed simply by waiting a year or two, or investing in larger systems. Clearly, an alternative approach is needed.

We present the concept of intelligent simulations as the answer to this call. In an intelligent simulation, we ask the system to work smarter, not harder. This can include replacing expensive parts of the model with a cheaper surrogate, or iterating through batches of simulations in such a way that minimal redundant information is determined.

In the news

“Applying machine learning at the front end of HPC”
23 Nov 2018, The Next Platform.
“IBM’s AI-HPC combine for ‘Intelligent simulation’: Eliminating the unnecessary”
16 Nov 2018, HPC Wire.
“Fueling the HPC transformation with AI”
12 Nov 2018, IBM THINK blog.

Intelligent experiments

AI in the lab

Automation is poised to revolutionize how scientific and engineering investigations are conducted both in the physical world, for example through robotic screening, and in the virtual world through modelling and simulation. Owing to rapid advances in relevant technologies, it is tempting — and somewhat risky — to focus solely on increasing the speed at which these experiments can be undertaken. It is at least as important, and arguably even more so, to improve the intelligence with which these automated investigations are performed. In order to do this, it is necessary to bring together knowledge, computation and generative technologies to guide these experiments in an intelligent manner. Each of these technologies individually has been shown to have the potential to accelerate discovery and innovation. However, no framework exists that integrates these three disciplines. For such a solution to work in an industrial setting, a solution is required to be algorithmically robust over a range of tasks, be able to be evaluated in parallel, and to scale to highly complex problems. In addition, it is necessary for the solution to allow itself to be operated by non-experts, which demonstrates a separation of concerns between the developer and the consumer.

Intelligent control

AI in the wild

Many infrastructure components are still controlled manually or with basic optimisation procedures. This can lead to highly inefficient operating procedures, which are often set to overly conservative configurations based upon worst-case scenario planning. Techniques such as reinforcement learning give us powerful tools to alleviate this problem and can produce a more scalable and sustainable solution.

One example is the treatment of waste water. Wastewater treatment plants are an important element of our water systems. In the UK, there are approximately 9000 wastewater treatment plants, and the sector accounts for around 20% of the UK’s annual energy expenditures. The current impediments to operate the system efficiently are highly variable wastewater influents due to variability in wastewater discharge and weather patterns. We are working on using reinforcement learning to operate the system under variable influents and reduce the overall cost of energy and chemical consumption, including potential violation fines.

3fundamental questions

3main technologies

See

How does AI adapt the way it views the world through data?

Think

How does AI make decisions and balance risk and reward when searching for new solutions?

Dream

How does AI take what it knows and uses it to construct new — as yet unseen — suggestions?

In order to achieve this, we work on three main technologies:

Deep learning. The construction of neural networks to build end-to-end learning machines capable of taking raw, unprocessed data, building representations, and using these representations to infer potential target values.

Bayesian optimisation. A black-box optimisation strategy that uses a probabilistic model to balance, explore and exploit complex search problems. One example of this is constructing the architectures of complex deep-learning machines.

Reinforcement learning. Coupling deep learning with traditional reinforcement-learning techniques (known as Deep-Q, or double Deep-Q learning) to teach agents how to respond to situations or environments in an optimal way.

Case study

Cognitive treatment plant

The cognitive treatment plant project aims to increase the efficiency of the water treatment process through the use of an artificial intelligence (AI) technique known as deep reinforcement learning. We teach the AI to minimise energy usage and disposal cost, while maintaining high quality effluent and minimise the number of regulatory violations. By exposing the AI to a wide variety of conditions, through a simulator we aim to ensure the robustness and stability of the control system. Common problems for operating a wastewater plant efficiently are that sensors do not always accurately measure the process; there are uncertainty in loads and volume of influents (which are dependent on weather and domestic/industrial usages).

As a result, wastewater plants are often run in a risk-averse way. For example the plant is operated as if a storm were coming, so more oxygen is pumped into the aerated tanks to ensure that the wastewater is treated properly even under adverse conditions. This means that there might be opportunities for smart pumping scheduling or reduced pumping to lower energy/electricity/operational costs based on using weather forecasts and a thorough analysis of which amount of pumping should be sufficient.

We have implemented AI on a controller that controls the pumping rate into an aerated tank. While in operation, the controller learns how different setpoints can lead to different influent qualities and energy costs. It then continuously learns and determines how to optimise the setpoints under different states of the tank so that it can minimise the operational costs while still ensuring that the effluents satisfy the treatment standard.