Partial Grounding in Classical Planning
||Partial Grounding in Classical Planning|
||NAISS Medium Compute|
||Daniel Gnad <email@example.com>|
||2023-06-29 – 2024-07-01|
The project is located in the area of sequential decision making in Artificial Intelligence (AI). Sequential decision making is one of the core challenges in AI. Developing systems that can achieve complex goals autonomously requires the capability to take a sequence of decisions that eventually lead the agent to its goal. AI planning has focused on enabling such capabilities for many years, but, in practice, existing solutions often fail to scale up to larger problems. In this project, we leverage the recent successes of data-driven approaches to overcome this limitation.
The project will focus on different aspects of the usage of machine learning for partial grounding. We consider four objectives that are partly connected to each other.
First, (1) we will explore different variants of data representation as input to the ML models. We plan to analyze the expressiveness of simple logic features that are used to describe task properties and whose evaluation is fed into an ML model. Furthermore, we will experiment with graph structures as well as textual inputs.
The second objective (2) targets the expressive power of different ML techniques. We will start to experiment with traditional ML models, such as decision trees and support vector machines. Leveraging recent successes of deep learning architectures, we will then explore the possibilities of neural network models, including graph neural networks and transformers.
Third, (3) we look into possibilities to make the overall planner more dependable. As partially grounded models come with no guarantees, e.g. might be unsolvable although the original task is not, we need a way to measure the robustness of the partial models. Moreover, the question arises what we can learn from a failed second phase if no solution to the partial model exists.
Lastly, (4) while the approach we take is generally tailored to learn ML models for a specific domain of planning tasks, i.e., a family of tasks framed in the same environment and dynamics, we want to extend our framework to enable learning for arbitrary planning tasks, across domains.
The project team consists of the PI and several PhD students (two are already added as members, some others will follow). All members will regularly conduct large-scale evaluations on the requested compute infrastructure. This will serve the purpose of assessing newly developed algorithms on a large existing benchmark set.