What Is Artificial Intelligence, Really?

06/07/2017 13:48

In popular media, "Artificial Intelligence" is by turns godlike, monstrous, uncannily human and a hoax; it inspires both awe and deep suspicion - it's unnatural.

Researchers who actually develop AI technologies - like those at - prefer narrower, more useful terms like Machine Learning (ML) and decision theory. They're wary of the catch-all phrase "Artificial Intelligence", in part because human intelligence is itself largely artificial, an encoded system of man-made concepts, rules of thumb, recipes, customs, laws, even whole cultures. Humans have always used thinking tools, rules and systems to keep chaos at bay. Turn off the traffic lights in central London and you'll soon see how far "natural" intelligence gets us in a complex system.

Following suit, AI has traditionally made decisions using painstakingly coded "if x then y" rules that sometimes appear intelligent. This works well in narrow, predictable, static environments like relatively simple games and machines but not in big, surprising, dynamic ones like cities, where trying to dictate every decision is madness.

The smartest complex systems are in fact made of well-coordinated autonomous individuals making their own decisions. That's how bee colonies and free societies work. In Machine Learning, those individuals are "agents": statistical entities that operate intelligently within computer models of environments like games, self-driving cars, and smart cities.'s agents get their smarts from three core technologies:

Probabilistic models:

Powerful statistical tools can generate flexible models of virtual or physical environments. Agents operate both in and on those models, effectively programming themselves and updating the models as they go along. No model is perfect; uncertainties and hidden relationships abound. One powerful statistical model, Gaussian Processes, can help estimate, account for and even reduce uncertainty, allowing the system to uncover hidden relations between events.

Reinforcement Learning (RL):

Agents can learn by acting in useful ways that are then reinforced numerically, much as a dog learns to sit when rewarded (reinforced) with a treat. Over time, the agent teaches itself to imitate, plan and perform sequences of actions, all without being given explicit instructions.

Multi-agent Systems (MAS):

Agents can also cooperate and compete using strategies adapted from game theory that benefit both themselves and the system as a whole. This helps them infer what other agents are doing and adjust for the often surprising, irrational behaviour of humans. The result is a safe, efficient, multi-agent system that is smarter than the sum of its parts.
The possible applications of machine decision making are virtually limitless, but let's focus on three examples:


ML will soon tackle the thorniest problem in gaming: maintaining player interest. The key here is offering an optimal level of challenge between getting bored when the game is too easy and frustrated when it's too hard. The next generation of ML will open up whole new classes of games with dynamic, evolving characters and storylines that that adjust to each player's style of play and provide personalized interactions. Really smart zombies, anyone? Development costs and time to market will plummet when testing is handled by teams of humans working with agents, who'll do the boring, repetitive jobs a thousand times faster than manual testers.

Autonomous Vehicles:

Get used to it, self-driving cars will increasingly take over our roads. Jaguar Land Rover is already testing a vehicle that is "nearly self driving" in city conditions. "If x then y" rules are a non-starter here: you can't program or script a vehicle to avoid ice patches, stray dogs or pedestrians. Put simply, probabilistic models will help a car "understand" itself and its environment, reinforcement learning will teach it to drive, and multi-agent systems will ensure it safely shares the road with other drivers, human and AI. Just as in gaming, ML can provide simulated environments where new technologies can be safely trained, tested and examined by regulators.

Smart cities:

Our increasingly complex cities need to get a lot smarter. ML systems will help regulators identify weak points like terror targets or fire hazards and ensure first responders intervene promptly. Well before construction begins on projects like the new runway at Heathrow, ML driven simulations will help planners design and test changes to infrastructure while taking into account the impacts of weather, pollution, people and vehicle traffic.

All this is but a small glimpse of the foreseeable future of Machine learning. It's the next few steps in a history of human intelligence that's always been driven by artificial information, technology and culture, by what we create as much as by what we are.