Principled AI:

Artificial Intelligence, Probabilistic Machine Learning and Principled Decision Making

back to our blogs

Principled AI:

Artificial Intelligence, Probabilistic Machine Learning and Principled Decision Making

Most people have an intuitive understanding of what “intelligence” means, but it's difficult to define formally. At we’re improving decision-making in complex systems, and that goal deeply informs our focus on intelligence as a pragmatic, practical concept, even if its exact boundaries are sometimes vague. For us, intelligence is characterised by autonomy in learning and decision-making in order to achieve goals. We use the term principled AI to describe systems with these properties built on coherent mathematical principles.

Natural and artificial intelligent systems excel at different types of skills. Even after years of dedicated training, humans are no match for computers in terms of raw precision, speed and memory. Conversely, no robot can compete with the phenomenal, seemingly effortless social skills of humans, which are the foundations of our societies. Attempting to replicate some of natural intelligence’s amazing abilities is a laudable goal and a future AI may eventually succeed in passing the Turing test. But we must also target challenges where human thinking reaches its limits, often because of the sheer amount of data to be speedily and dynamically processed.

Decision making is crucial — if your adaptable information processing system isn't going to make or inform decisions, then you may as well not bother. If you want it to make useful, efficient decisions, then you’ll need to design and integrate the decision-making process into your system from the start, not just tack a decision module onto a separately designed learning system. As artificial systems get larger, more complex and more sophisticated, manual engineering and design become increasingly intractable. Instead, future AI systems need to be based on automated inference within a principled framework. Fortunately, probability and information theory together with Bayesian decision theory provide a mathematically principled and practical approach to solving these problems.

Building successful AI systems requires three pillars of expertise, corresponding to our three technical research groups at

Probabilistic Modelling

Adaptable, learning systems are driven by data. But because data is inherently limited in quantity and often noisy, the learnt models will always be subject to some level of uncertainty. Probability theory provides a coherent framework for reasoning about uncertain events which is ideally suited to this task. James Clerk Maxwell knew this already in 1850:

"the true logic for this world is the calculus of Probabilities, which takes account of the magnitude of the probability which is, or ought to be, in a reasonable man’s mind."

Probability theory is the foundation of information theory, and provides a practical framework for reasoning and learning by Bayesian inference. It provides automatic tools for model fitting, prediction, model comparison, etc.

Reinforcement Learning

Decision making is married to machine learning through Reinforcement Learning (RL). In RL an agent takes actions and receives information about the state of the world as well as rewards, indicating to what extent the desired goals are being achieved. The task of an agent is to use all available information to figure out which actions to take. The behaviour of the world is typically not deterministic and the learnt models have uncertainties, so RL needs probabilistic approaches to make good decisions. The best decisions depend not only on the most likely predicted outcome, but also on the confidence associated with the prediction and on the consequences of possible outcomes. Well established theory including Bayesian decision theory and Bellman optimality, form the basis of RL and meet the requirements for intelligent behaviour.

Multi-agent Systems

A final important component for intelligence is multiple agents. Realistic scenarios in which agents can consider their task or environment static are rare. Usually environments dynamically change as agents react to each other’s actions. An intelligent agent must model the behaviour of other agents in order to understand their intentions and why they react the way they do. The mathematical discipline of game theory formalises these ideas, and sheds light on the strategic behaviour of multi-agent systems, including exploitation and cooperation.

The Principled AI Company

These three pillars of principled AI form the basis for solving problems across a staggering range of applications and industries. Whether optimising supply chains in the face of uncertain demands or constraints, controlling robots, making financial decisions, understanding user actions, or optimising fleets of agents, the underlying principles remain the same: probabilistic models, automated inference, and goal-directed fully integrated mathematically-principled decision making.

That’s what we do at

Join us to make AI that will change the world

join our team