Although AI Planning and Constraint Programming share many techniques and approaches, an important difference lies in the approach to modelling. In CP and also in Operations Research, modellers spend considerable time and effort evaluating alternative models and selecting representations of a problem that will make it most amenable to solution by existing technology. In Planning, researchers typically spend little time considering alternative models and are content to work with the first model they construct, working instead on improving the planning technology to try to tackle the problem, whatever its form. The reason for the strategy of planning researchers is that the intention is to avoid the need for expert planning knowledge in order to exploit a planner. However, the price for this strategy is that there is very little accumulated research expertise in the problem of modelling and no systematic comparison of the performance of planners using alternative models of the same problem. Although avoiding the need for expert planning knowledge in order to use a planner is an important goal, there is clearly a lost opportunity to identify ways in which models might be structured to be most amenable to solution. We propose to combine these strategies by exploring the automatic reformulation of planning problems in order to better exploit the existing planning technology by restructuring models to expose the information that can make a planner make more intelligent choices.
Our group was the first to start extending forward search planners towards solution of problems with continuous time and numeric quantities. Over a number of years, and funded by a sequence of projects, this has led to a very capable planning framework, called POPF, which is still leading the field in temporal and metric planning in mixed discrete-continuous domains.
|Effective start/end date||1/01/09 → 30/11/12|
- EPSRC (Engineering and Physical Sciences Research Council): £343,967.00