Object-Oriented Transition Modeling with Inductive Logic Programming
Gabriel Stella, Dmitri Loguinov
TL;DR
This paper introduces a new object-oriented transition modeling algorithm using inductive logic programming, achieving superior accuracy, interpretability, and efficiency compared to previous methods and neural baselines.
Contribution
The paper presents a novel, more powerful learning algorithm for object-oriented transition modeling that outperforms existing approaches in accuracy and efficiency.
Findings
Significant improvement over state-of-the-art methods
Demonstrated robustness through ablation tests
Outperformed neural baseline models
Abstract
Building models of the world from observation, i.e., induction, is one of the major challenges in machine learning. In order to be useful, models need to maintain accuracy when used in novel situations, i.e., generalize. In addition, they should be easy to interpret and efficient to train. Prior work has investigated these concepts in the context of object-oriented representations inspired by human cognition. In this paper, we develop a novel learning algorithm that is substantially more powerful than these previous methods. Our thorough experiments, including ablation tests and comparison with neural baselines, demonstrate a significant improvement over the state-of-the-art.
Peer Reviews
Decision·Submitted to ICLR 2026
- The approach performs much better than prior work. It can handle variable bindings, nested quantifiers, and more complex rules. - The algorithm design seems reasonable, although many details of the implementation are not described well.
Two main weaknesses: impact/significance, and lack of clarity describing the main algorithm. W1: impact and significance - Object oriented transition modeling is a domain with limited impact. This is good old fashioned AI, with symbolic observations, learning, and models. It's unclear to me what the insights from 2025 are compared with previous decades of research into this area. - The evaluation is not very thorough: only one baseline is compared to. The environments are small gridworld env
The paper conducts several ablation experiments to validate the effectiveness of the proposed method.
Although the paper states its effectiveness, I believe it lacks adequate discussion of the performance differences among the various baselines. Additionally, some experiments still lack key comparative results. Furthermore, the presentation of this paper still needs to be improved. For example, the paper does not include a discussion of the preliminaries or related work, which would be useful for contextualizing the state-of-the-art models within the inductive logic programming and reinforcement
- Interpretable model class learning (FOLDT) - Object centric nature
- Method is described too abstractly in the main text; intuition and operational details are hard to extract - Strong modeling assumption: attributes are predicted independently; the paper does not justify, analyze, or ablate this restriction. - Reward modeling: including cumulative reward as part of state breaks the standard MDP Markov assumption unless extra conditions hold; this is neither acknowledged nor analyzed. - related work is thin given a long history in action-model learning, model-b
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsLogic, Reasoning, and Knowledge · Topic Modeling · Data Mining Algorithms and Applications