Autonomous emissions detection through probabilistic decision making

We're increasingly using autonomous vehicles for a range of operations. As we do this, the period of expected operation is increasing and the level of required autonomy will follow suit. In many of these scenarios we might not know what the optimum behaviour will be before we start the operation.

So the vehicles will need to balance the time they spend improving their performance and exploiting the knowledge that they've gained. An example of this is determining which vessels are breaking legislation around emissions within UK sovereign waters. We need to be able to police these vessels and determine their levels of NOx and SOx emissions.

One option is to send a drone but these have a limited range and must decide how they can act most effectively in each mission, depending on what is observed: the drone can try to examine as many vessels as possible, gather data to improve its future predictions or target those it thinks are most likely to be polluters.

It will need to optimise its behaviour to maximise the number of positive tests over time, rather than looking for short term benefits. This type of scenario is prevalent in other applications, such as searching for gas leaks or mapping the seabed.

This PhD will generate an approach to AI that allows it to adaptively change its performance over time, within the constraints of a low power, low compute autonomous vehicle.