Pierre Haren is the co-founder of ILOG and an expert systems specialist. He contribued to the birth and the rise of what is now known as artificial intelligence. He tells us about about his career and shares a few thoughts about the future.
Polytechnique, Ponts et Chaussées and MIT
Expert systems researcher at Inria
CEO of ILOG
Worldwide Leader of Advanced Analytics chez IBM
Co-founder and CEO of Causality Link
Birth of Artificial Intelligence
After l’École Polytechnique, l’École des Ponts, and MIT, I went on to work as a research administrator for the French Ministry of the Sea until 1983. However I wanted to get into artificial intelligence, which I had learned about while doing my thesis at MIT in the late 1970s. At the time, there were two institutes that I could work with: the Centre Mondial Informatique et Ressource Humaine founded by Jean-Jacques Servan-Schreiber in 1981, but which closed in 1986, and Inria. Gilles Kahn, who was Scientific Director of the Inria centre in Sophia-Antipolis, and Pierre Bernhard, Director of the centre, were intrigued yet sceptical about letting an engineer from l’École des Ponts join their research. They put me to the test for a year in 1983! It was exciting because the centre in Sophia-Antipolis was in its infancy and I was earning my stripes at an institution that was also just starting out. When the King of Spain came to visit to purchase helicopters from Aérospatiale, we had to bring down all the chairs from the second floor to make the ground floor look fuller! I had joined a unique institute that was on its way to developing its trailblazing status.
I then developed a project called SMECI, drawing on all the advantages the institute had to offer, such as autonomy, freedom and the scientific environment. I also learned a lot while I was there. My project was a success as, after four years, we had clients and an original prototype that demonstrated the potential of multi-expert systems in artificial intelligence. We couldn’t find anyone willing to manufacture the product so we did it ourselves by creating ILOG. This industrial adventure was the logical next step after our research. In any case, almost all the projects at Inria had concrete applications in the real world.
We stayed in contact with former research teams and worked a lot with people like Gérard Berry (for whom we produced and sold Esterel software at one time). Being your own boss is very different from working as a researcher at Inria. Both private and public clients don’t pay at the end of the month and we almost had to close up shop right from the get-go! This reality check forces you to change your priorities and, at some point or another, relationships inevitably grow thin between the more protected world of research and the world of business, which focuses on growing your client base and financial stability. Nevertheless, I still managed to continue as a scientific advisor to successive presidents of the institute over several years.
ILOG was a long adventure! The first ten years were spent establishing subsidiaries abroad and becoming a NASDAQ-listed company. We had major clients in France, such as Peugeot and Thomson-CSF (now Thales), but some companies, like BNP-Paribas and Dassault Aviation only began buying our products after ILOG was bought out by IBM. Major French corporations were often unwilling to bank on home-grown start-ups, so we expanded into Singapore, Spain, and the United States, which in the end, made up half our sales. We sold ILOG to IBM in 2008 and it became one of its five best acquisitions. They saw a return on their investment within three years. The transition was a success and sales shot up around the world, especially in France with BNP-Paribas! After eight years at IBM, I decided to get back into entrepreneurship with a new start-up, Causality Link , which specialises in Artificial Intelligence for finance.
From Lisp to expert systems
The first wave of artificial intelligence technology produced Lisp machines, bitmap visualizers and advanced graphical user interfaces. In the 1980s, we were in the second wave of AI, focused on expert systems. These were systems based on object models and rules that could emulate human experts. My project designed and developed a pioneering multi-expert system generator that could liaise several lines of reasoning, including a knowledge-based line of reasoning (Rose Dieng’s first subject of research) which adapted to the user. No one else had done it before. It was really the birth of what became known as knowledge engineering – modelling the way experts approach a project. One of our clients was the Ministry of the Sea, for which we developed a port engineering knowledge management system that was used for about ten years. We also created a simulation model for nuclear submarine pilots for the French Navy.
With ILOG, we then rolled out hundreds of expert systems, autonomous assistants in areas varying from decision simulation for naval battles to bank transaction validation, automobile transmission configuration and standards compliance management in the building industry to speed up drawings approval and construction times. We deployed the world’s first generation of automatic Vessel Traffic Management Information Systems. Our system monitored, for instance, the movements of all gas tankers, where accidents can have disastrous consequences, in the port of Saint-Nazaire, then Marseille and the training centre in Rotterdam. The technology was sold to Lockheed Martin, which became the global leader in automatic Vessel Traffic Management Information Systems. We were the first to replace radar displays with a computer display connected to an expert system.
Each new wave of computer and artificial intelligence technology created products that were useful for the general public. For instance, expert systems and rule-based systems produced technologies that have become standard tools. No-one considers them as artificial intelligence now because they have become part of everyday use.
IA for tomorrow ?
We’re no longer in the explicit AI era but in the third wave of implicit AI with neural networks, where we don’t know exactly what the machine has learned. Experts aren’t getting better by watching the machine operate either. In the 1980s, the primary purpose of a system that detailed the reasoning of a machine was to make it communicable, detect the bugs and improve it. I think that this type of artificial intelligence was unjustly undermined by the success of implicit AI. Today, advances are being made in closed worlds where a reasonable copy of reality can be simulated. Open worlds in which we don’t know how to simulate forms of reality, such as the world of finance, won’t make it without explicit AI.
My first digital memory
I started programming at l’École Polytechnique in 1973. Our grade averages over the first two years determined our class ranking and which professions we could access. The military personnel gave grades for each assignment but didn’t calculate the rankings after each new grade. We had to wait until the end of the two years. I took my punched Fortran cards and created a system to give the rankings in real time. There was a data positioning error in the card columns so my first ranking was completely wrong, which made all the military personnel and my friends freak out. So my first memory of digital technology was actually a bug!
I don’t know what the future holds for this third wave. Existing logical capacities are being imitated and extended more than they are being invented. Neural systems actually existed before expert systems, with Rosenblatt’s perceptrons. Current neural networks imitate the electrical rather than the chemical aspects of how the brain works. We’re still a long way off from the human brain. The next wave will perhaps be an extremely interesting combination of neural networks, predictive analytics models and existing knowledge in explicit artificial intelligence. When combined, fascinating things could be done! Today, mathematicians are developing their own formulas, cognitive scientists are explaining human reasoning and neural networks are absorbing huge numbers of situations and learning regularities. We can therefore hope that cooperation between these three branches will help experts learn from machines and correct them in open worlds. This works in closed worlds where all the rules are known, such as chess, poker and Go. However, learning from mathematical models in open worlds, such as driving, is still in its inception.
Artificial intelligence will be applied to spaces and situations that we are able to simulate and/or predict, from self-driving vehicles to medical simulation. There have been and will be huge advances in statistics, natural language and predictive analytics, which involve a lot of mathematics. In France, we’re on the cutting edge in the field.
More generally, we’re on the brink of creating a model of humanity. With Facebook, we went from modelling individuals to modelling a global network of individuals. Other human network models are already being used, such a consumer behaviour (Amazon) and behaviour in the world of employment (LinkedIn). These three human models, (friends, consumer behaviour, work relations) will converge to build a model (with big data) of human beings in society, which will be the matrix for future advances in computing. This will be followed by automation questions in transport, medicine, etc., that rely on human data.
I’m 63 years old and I recently created Causality Link , a business that specialises in financial artificial intelligence. We're working to combine implicit and explicit systems where machines will improve how experts in the field do their job, rather than replace them! So stay tuned!