Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Object-Oriented Transition Modeling with Inductive Logic Programming

Gabriel Stella, Dmitri Loguinov

TL;DR

This work tackles the problem of learning accurate, generalizable, and interpretable object-oriented transition models. It introduces TreeThink, an ILP-based framework built on TreeLearn and First-Order Logical Decision Trees (FOLDTs) to learn per-class, per-attribute, per-action transition rules that predict attribute deltas and compose into the state transition function $\hat{T}(s,a)$. Empirical results show TreeThink achieves zero error across diverse domains, outperforms the prior ILP method QORA and neural baselines, and benefits from inference optimizations, branch updating, and strong scaling behavior. The approach enables efficient online learning and planning in complex, reward-bearing environments, with code and extensions likely to advance planning and reinforcement learning in object-centered representations.

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.

Object-Oriented Transition Modeling with Inductive Logic Programming

TL;DR

This work tackles the problem of learning accurate, generalizable, and interpretable object-oriented transition models. It introduces TreeThink, an ILP-based framework built on TreeLearn and First-Order Logical Decision Trees (FOLDTs) to learn per-class, per-attribute, per-action transition rules that predict attribute deltas and compose into the state transition function . Empirical results show TreeThink achieves zero error across diverse domains, outperforms the prior ILP method QORA and neural baselines, and benefits from inference optimizations, branch updating, and strong scaling behavior. The approach enables efficient online learning and planning in complex, reward-bearing environments, with code and extensions likely to advance planning and reinforcement learning in object-centered representations.

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.
Paper Structure (43 sections, 4 equations, 26 figures, 1 table)

This paper contains 43 sections, 4 equations, 26 figures, 1 table.

Table of Contents

  1. Introduction
  2. TreeThink
  3. First-Order Logical Decision Trees
  4. TreeLearn
  5. Updating Tests
  6. Updating Nodes
  7. Recursion
  8. Inference Optimizations
  9. Experiments
  10. Environments
  11. Comparison with Prior Work
  12. Ablation and Performance Tests
  13. Inference Optimizations
  14. Branch Updating
  15. Hyperparameter Robustness
  16. ...and 28 more sections

Figures (26)

  • Figure 1: An example game transition with a partial listing of the initial state in the object-oriented representation. Although we use meaningful names to increase readability, the information given to an agent does not contain these labels.
  • Figure 2: FOLDT and equivalent program representation of a rule for the keys domain.
  • Figure 3: TreeThink vs. QORA predictive modeling
  • Figure 4: Ablation and Performance Tests
  • Figure 5: Using various NPE architectures to model the maze domain. Semilog-$y$ plots are used to better visualize small values. Breaks in a line (e.g., the error of NPE_8_3) indicate zeroes.
  • ...and 21 more figures