Neurosymbolic world models improve interpretable planning over long horizons
Neurosymbolic Learning of Interpretable World Models
Artificial Intelligence
Summary
Predicting what will happen next in a game or environment is hard because computers usually learn in ways that are hard to understand. The authors created a method that uses clear, human-readable rules together with learning from experience to build concepts the computer can understand. This method helps computers better predict future events and make plans that work well, even when things change. They tested it in games and found it works better than purely neural methods or simpler neurosymbolic ones.
What this means in practice
- •For robotics developers: Build robots that interpret and plan actions based on understandable concepts derived from formal knowledge and observations.
- •For game ai programmers: Create game agents that use explicit logical rules combined with learning to predict game states and make reliable plans over long time spans.
Authors
Riccardo Andreoni, Massimiliano Ronzani, Chiara Ghidini, Alessandro Daniele, Emile van Krieken
Abstract
World models aim to learn compact representations of an environment from observations, enabling agents to predict future states and plan their behaviour. However, most existing approaches learn black-box representations, with limited ability to incorporate prior knowledge about the world. We introduce Concept-Grounded World Models (CGWMs), a neurosymbolic world model design that instead learns interpretable concept representations. We learn these concepts from observations using explicit and complete knowledge, expressed in temporal logic, that specifies the valid world states and the temporal relations governing their dynamics. Furthermore, we directly use this knowledge by formulating planning as differentiable reasoning, optimising the agent's actions by gradient ascent to satisfy a goal. We experimentally compare CGWM with a purely neural baseline and a neurosymbolic baseline on two game environments. We find that CGWMs ground concepts much more accurately, even over long horizons, resulting in significantly more robust planning both in and out of distribution.