State of the Art:
Research on autonomy includes the development of fundamental search‐driven techniques that support classical planning and the extension of these techniques to cover a wider range of expressivity, such as planning with time and numbers, continuous change and uncertainty.
This project will exploit the relevant state of the art, for example in action‐modelling, search control and exploitation of search topology features such as landmarks. Recent work in hindsight optimization will be exploited here. This offers a powerful alternative to planning with probabilistic models. In hindsight optimization, the finite‐horizon utility achievable from a given system state is estimated by averaging estimates obtained from a number of traces starting at the state. For each trace, the utility value of the state is estimated by determining the optimal hindsight action choice. The planner, that somehow “knew” the whole trace beforehand, measures the utility obtained under that action. Averaging over many samples gives a simulation‐based hindsight‐optimal utility for the starting state.
One of the most promising approaches to building uncertainty-robust plans uses hindsight optimization discussed above. The idea is to literatively improve a plan by considering samples drawn from the space of alternative roll‐outs of the execution of the current candidate plan. This approach can be supplemented by clever sampling strategies to focus on appropriate parts of the search space. These techniques provide a good baseline for the development of persistent autonomy in Pandora.
On top of this, we will:
- Allow the removal and insertion of plan fragments to modify a plan under execution.
- Support the insertion of plan fragments to use up released resources in ways that allow the system to take advantage of “opportunities”.
- Modify the plan while preserving high confidence that the plan will achieve its highest priority goals.
The computational mechanisms involved in achieving robust plan modifications are constraint reasoning and convolution of the distributions representing probable resource demands of actions.