Hybrid AI from the point of view of Arnaud de Moissac, DCbrain’s CEO

What would be a simple definition of Hybrid AI?

Hybrid AI can be defined by two complementary approaches at DCbrain. First, it is about injecting physical laws to an artificial intelligence in addition to data in order to anchor these AI in reality. Secondly, it is about giving a quasi human intuition capacity to the algorithms used to optimize networks in order to obtain and generate artificial intuitions. The Hybrid AI is therefore a “mix” of artificial intelligence, human expertise, theoretical behavior and real world data measurements.
This technique allows INES, DCbrain’s platform, to learn very quickly a very precise digital twin, which is limited by physical limitations even if few data are available. This digital twin will improve over time. It also allows to obtain extremely short optimization computation times on very complex problems.

An emblematic example of Hybrid AI is Google’s victory over the world Go champion. This victory did not seem within the reach of a computer for several decades. Google hybridized a classic Go game algorithm (MCTS) with neural networks and then allowed this algorithm to play several hundred thousand games against itself (this is called Reinforcement Learning) to acquire a kind of human intuition at the level of a great master.

In what context do you use Hybrid AI at DCbrain?

Our customers are facing a real revolution in the way they operate their networks. The injection and consideration of renewable energies, for example (such as biomethane or hydrogen), means that networks that were previously essentially managed on a medium- or long-term basis have to operate under constraints on a daily basis. Operators find themselves having to manage new situations with many hazards every day. Yet their traditional tools are unable to take advantage of the data coming from the network to help them in this new task.

At DCbrain, Hybrid AI is at the heart of our solution for gas, energy, water, electricity, heat, transport, logistics, etc… It allows our customers to optimize their networks, visualize them, simulate different scenarios, detect anomalies and make forecasts. All this is possible with a very short learning time and a particularly accurate digital twin since it is based on field data.

What are the advantages of Hybrid AI compared to other techniques?

The Hybrid AI is a much faster approach, which means that many more scenarios can be tested and therefore an ideal optimization can be achieved.
For example, the construction of the digital twin for a gas network goes from 1 year for a classical solver to 1 month for INES. In addition, DCbrain’s digital twin is continuously improving. The calculation of the optimal configuration of a network based on context data (sensors, weather…) takes 1 day for a classical solver but only 1 minute for DCbrain.

Moreover, Hybrid AI provides the same power as Artificial Intelligence while avoiding the black box effect, i.e. the misunderstanding of its reasoning. Indeed, as Hybrid AI uses physical laws, it is possible to understand its analyses. We talk about explicability. Even if Artificial Intelligence has a lot of data at its disposal, it remains prisoner of what it has learned. Thus, without physical laws, it has its limits and cannot calculate what deviates too much from what it has learned. Finally, the Hybrid AI is easier for professionals to understand because it can display network mapping in real time, which would allow the detection of anomalies quickly for example.