Entry for the EPSRC Computer Science Writing Competition 2007
by Tom Stafford
You track the ball as it begins to fall towards you. Nelson stands between you and the goal. Before the ball lands you tap it with your foot for control, feign right but then pause - because Nelson knows you well enough to expect a feign. At the pause he goes left and you drop right, slamming the ball into the back of the net. You turn to where you know your dad is in the crowd, smiling.
You're late, heading down the M1, trying to find something to listen to on the radio, but watching that blue Corsa in the inside lane - something about the way it's been moving makes you nervous. You're twiddling the dial, and, - yes! - you've found some Motown, but - look out! - the blue Corsa swerves into your lane without indicating. It's okay, you apply the brake and ease back out of trouble.
Your youngest is trying to tell you something about sharks while you are pulling his coat over his head with one hand, holding the mobile phone between your head and shoulders to talk to Daphne about where you want to go on holiday. With your other hand you lock the door.
Doing one thing at a time is hard, coordinating doing several things at once is a work of genius - but our brains to it for us constantly. Are there general principles which can be used to solve this 'action selection' problem? Kevin Gurney, principle investigator of the EPSRC's REVERB project, thinks there are.
"We believe", says Dr Gurney, "that the fundamental organisation principles which allow us to coordinate multiple behaviours are embodied in the architecture of the brain - particularly in the basic subcortical architecture which all vertebrates share"
The REVERB project is using a combination of computer science and neuroscience to discover what these principles are. REVERB researchers in labs across the UK are working on the construction of a robot which uses computational principles derived from experiments on humans and animals.
Dr Tony Prescott, of Sheffield University, explains the philosophy of the approach. "Psychology and Neuroscience are both what I call 'analytic'. They look at a system - the brain - and try and deduce how it works. Computational modelling and robotics are 'synthetic'. They look at how a set of principles can be used to solve problems, either in a virtual world or, with robots, in the real-world."
The team is convinced that computer science gives a conceptual foundation to neuroscientific investigation. "Without looking at functional principles - 'computations', in other words - you're just lost in a sea of biological detail.", says Professor Peter Redgrave, head neuroscientist with the REVERB project. So what does neuroscience offer computer science? "In the vertebrate brain we've got an expert system, refined over millions of years by evolution, and at heart basically preserved in the same form across all animals, which could be the biological solution to the problem of how to prioritise and coordinate actions".
It's not a trivial problem, as the above examples illustrate. You have to start and stop actions at the right times, trying to make efficient compromises where you can (like holding your keys with one hand, and your youngest child with another). But in some situations compromises between actions just aren't appropriate, you have to quickly do one thing or the other (if you need to dodge left or right, sticking in the middle is worse than either choice). Added to this is the difficulty of incorporating different motivations at different levels of abstraction (how do you trade-off the desire to find a good radio station with the need to stay safe on the road? Combine what you know of a particular defender with the physical information about a ball's trajectory?).
If the REVERB project can unravel these secrets of behavioural integration from the vertebrate brain then it can make progress on building an autonomous robot - an agent that can adaptively explore its environment and react to what it finds, even learning from its actions along the way.
The first goal of the project is a robot what can visually scan its environment, making sense of what it finds there and focussing attention on important objects and then moving appropriately - just as humans do when they move their eyes over a scene. This will involve all the generic problems of behavioural coordination that the REVERB project is setting out to solve.
"The goal is to reverse-engineer the computational principles that the brain uses to multi-task", says Dr Gurney, "and to show that they work in practice, not just in theory, by incorporating them in a robot. In a sense, we're trying to work out the real-time operating system of our own minds."