-
Natural sciences
- Machine learning and decision making
In this project we investigate the potential of modern generative models in the context of planning and decision problems. For this problem family, it is important to analyse scenarios of possible futures in terms of task fulfillment. The presence of uncertainty (caused by the stochastic nature of the environment) as well as inaccuracies (caused by only partial observability of the environment) is an important complication.
We propose modern generative models as an alternative for the current practice of combining offline trained recurrent neural network world models in combination with Monte-Carle Tree Sampling. Specific attention will be directed towards diffusion models, tensor networks and transformers. The formulation of planning problems in terms of these generative models, originally designed for generation of images and text, is a major challenge.