Learning General Policies for Planning using GNNs and Transformers
||Learning General Policies for Planning using GNNs and Transformers|
||Markus Fritzsche <email@example.com>|
||2022-11-10 – 2023-06-01|
In classical planning, it is desirable to learn general policies from samples of input states.
These usually consist of a set of true atoms, but in real world applications they are encoded as raw perceptions, e.g. in games as images. One approach to learning general policies from these perceptions is either to learn the associated true atoms or to learn a policy directly from the input perceptions. We believe that Deep Learning methods excel for several reasons. (1) Recent work has shown that it is indeed possible to learn general policies from STRIPS state representations. (2) The state-of-the-art in various computer vision tasks, e.g., object detection, has been based on Deep Learning for many years. Therefore, we want to investigate whether it is possible to learn general policies from raw perceptions by using neural networks directly or by transforming perceptions into a STRIPS state representation. More precisely, we want to focus on Graph Neural Networks (GNN) and Transformer models to train greedy search heuristics.