Relational models predict state changes for better action planning
Relational Learning as World Modeling
Machine LearningArtificial Intelligence
Summary
Most relational learning focuses on guessing the current state of a system, but this paper tackles predicting how that state changes after an action. The authors created a model called TabWM that learns from observing how different systems behave over time without needing to adjust its parameters. This model can then predict detailed changes that happen after new actions, allowing it to imagine future states and evaluate possible next steps. Their tests showed that seeing more example runs improved prediction accuracy a lot, helping choose better action sequences. This approach could help machines better understand and interact with complex relational environments.
What this means in practice
- •For robotics developers: Predict how robot actions change their environment to plan better task sequences without retraining models each time.
- •For game ai programmers: Simulate possible game world changes from player actions to improve non-player character decision-making and strategy planning.
Authors
Zheyu Zhang, Gjergji Kasneci
Abstract
Relational learning typically predicts from a database state, but decision making asks a different question: how will that state change under an action? In operational systems, the answer depends on system-specific dynamics that may not be visible in the database itself. We view each such system as a relational "world" and ask whether a model can infer how it behaves from observed executions. We introduce TabWM (Tabular World Model) and WorldRel, a generator of executable relational worlds for pretraining. Given executions from an unseen world, TabWM uses them as context to predict the sparse structured edits caused by a new action without parameter updates. Across 512 held-out worlds, complete-edit F1 rises from 0.016 with no observed executions to 0.685 with 64, where a correct edit must match its operation, relational location, and value. Because these predictions are explicit relational edits, they can be applied recursively to imagine candidate futures and compare actions before they are taken. Broader search over these predicted futures improves the realized return of selected action sequences. Together, these results suggest that relational dynamics can be learned in context, broadening relational learning from modeling system state to modeling how that state changes under action.