Abstract:Deterministic planning assumes that the planning evolves along a fully predictable path, and therefore it loses the practical value in most real projections. A more realistic view is that planning ought to take into consideration partial observability beforehand and aim for a more flexible and robust solution. What is more significant, it is inevitable that the quality of plan varies dramatically in the partially observable environment. In this paper we propose a probabilistic contingent Hierarchical Task Network (HTN) planner, named High-Quality Contingent Planner (HQCP), to generate high-quality plans in the partially observable environment. The formalisms in HTN planning are extended into partial observability and are evaluated regarding the cost. Next, we explore a novel heuristic for high-quality plans and develop the integrated planning algorithm. Finally, an empirical study verifies the effectiveness and efficiency of the planner both in probabilistic contingent planning and for obtaining high-quality plans.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the challenge of generating high - quality plans in partially observable environments. Specifically, classical deterministic planning assumes that the planning process develops along a fully predictable path, which loses its practicality in most real - world scenarios. A more realistic situation is that planning should take partial observability into account in advance and seek more flexible and robust solutions. However, in partially observable environments, the quality of plans can vary significantly due to uncertainty, so simply finding feasible plans cannot meet the requirements.
To this end, the authors propose a probabilistic conditional planning method based on Hierarchical Task Network (HTN), called High - Quality Contingent Planner (HQCP), to generate high - quality plans in partially observable environments. The following is a specific description of this problem:
1. **Uncertainty in Partially Observable Environments**: In partially observable environments, there is uncertainty in the initial state. It is necessary to obtain real - world observations through sensing actions and generate conditional plans based on different observations.
2. **The Need for High - Quality Plans**: In uncertain environments, the costs of different plans vary greatly, so it is necessary to find high - quality plans with the minimum cost. For example, in fire emergency evacuation, the specific locations of multiple trapped people are unknown, but all people need to be evacuated at the minimum cost.
3. **Limitations of Existing Methods**: Although existing HTN planning methods can handle partial observability or generate high - quality plans, they have not combined the two, that is, research on generating high - quality plans in partially observable environments is still blank.
To solve these problems, the paper makes the following contributions:
- **Extension of HTN Formalism**: Extend the formalism of HTN planning to partially observable environments and introduce concepts such as belief states and sensing actions.
- **Cost Evaluation**: Define how to quantitatively evaluate the quality of plans, especially to distinguish plans of different qualities by cost.
- **Heuristic Algorithm**: Explore a new heuristic method to guide the search for high - quality plans.
- **Integrated Planning Algorithm**: Develop an integrated planning algorithm HQCP that combines probabilistic conditional planning and heuristics.
- **Experimental Verification**: Through empirical research in two partially observable domains, verify the effectiveness and efficiency of HQCP, including the Medicate domain and the extended ZenoTravel domain.
In summary, this paper aims to fill the gaps in existing research, provide a new method for generating high - quality plans in partially observable environments, and verify its effectiveness and superiority through experiments.