In recent years planning technology has a enjoyed significant increase in real-world application, with industry and general science research benefiting from the great depth theoretical work done in this area over many decades. The core problem of deciding which activities to carry out and when occurs in a vast range of domains; the research area of planning is concerned with developing generic problem solving technology that automates the task of performing this core reasoning. Planning technology has been employed in a wide range of domains including controlling printing presses, power management for the UK National, train scheduling on the Hungarian Railway Network, scheduling aircraft landing in airports, and autonomous robotic control, both in space and in the oceans. Experience in these areas has given rise to two key observations. First, the existing theoretical work done in AI Planning has been extremely valuable, allowing planning technology to begin to solve real world problems. Planning is a fundamental component of intelligent autonomous behaviour and as such planning technology has real potential for application in many different areas, both now and in the future. The second is that whilst one can observe that planners can now begin to be applied to these problems, there is still a great need for improvement of the underlying technology, in terms of expressivity and performance, in order to be able to create greater autonomy by allowing reasoning about an uncertain world. At the heart of this lies deeply theoretical computer science research: a planner is a generic problem solving system, consisting of search algorithms and heuristics. Of particular interest is reasoning about time and resources, something key to many areas of computer science, from compilers and programming languages to web services and optimisation. In order to tackle application problems well, reasoning effectively about these is essential. Of specific interest here is uncertainty in time taken and resources consumed. This occurs in many application domains, and in each of these a similar approach is taken: conservatism about time and resource availability in order to guarantee success. This, however, comes at a cost. By way of example, when planning for autonomous Martian exploration, the models used by both the ESA and NASA are pessimistic, underestimating the amount of power the rover will receive from the sun, and overestimating the amount of energy and time each activity will take. The result is that the equipment is highly under-utilised, with fewer science targets being achieved than could have been with better on-board reasoning. Given the expense of placing rovers on Mars and the limited equipment lifespan, this is a great cost to mankind's exploration of space. A similar problem occurs when deploying renewable energy generation: wind farms are assumed to provide 10% of their maximum output, even thought the reality is almost always greater than this. This conservative assumption, there to ensure power is always provided, causes great environmental and economic cost, as extra production capacity must be available through other sources regardless of whether it is required. The major benefit of planning is in generating a generic problem solving technology. Developing several bespoke solvers would take many years, and incur great financial cost. By developing efficient planning systems, a single domain-independent problem-solving core is built, capable of solving many problems without the cost of developing a bespoke solver for each. The core of this research is addressing challenges in solving the general planning problem that will allow future application of planning, extending the range of problems to which this generic technology can be applied.
|Effective start/end date||20/04/10 → 30/04/13|
- EPSRC (Engineering and Physical Sciences Research Council): £203,745.00