Abstract
Uncertainty hinders many applications of AI { algorithms are often faced with noisy sensors, unpredictable environments, or known limitations in world models.Automated planning with numeric uncertainty is an active research field that can tackle some of these challenges already. However, areas such as computing heuristics and building policies still pose open questions.
In this thesis, we aim to improve planner performance on domains with numeric uncertainty. To this end, we implement several techniques into a novel planner based on the existing Optic+. Our contributions are connected by an over-arching theme: tracking uncertainty and integrating it with aspects of planning that previously ignored it.
We begin by improving heuristic guidance for forwards planning with continuous random variables, by ensuring preconditions are met with a given degree of confidence. With this new information, the planner can also consider acting to reduce accumulated error. We then go on to define regression semantics for the case where uncertainty is Gaussian, allowing forward-policy building to be extended to handle not just propositional uncertainty but also numeric uncertainty. We also propose an internal representation that allows the planner to sample action effects from any probability distribution. As a consequence, the world model can be refined without sacrificing computational time by adding more samples during execution.
While our original motivation { and our running example throughout this thesis { is robotic exploration, our contributions are general and not limited to one type of problem. Ultimately, we reduce the effort needed to describe uncertain domains and show how planners can tackle a changing environment while meeting given certainty requirements.
| Date of Award | 1 Jun 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Amanda Coles (Supervisor) & Andrew Coles (Supervisor) |