Learning Integrated Symbolic and Continuous Action Models for Continuous Domains

Joseph Zhen Ying Xu, John E. LairdAAAI 2013 · DOI

Introduction

In the field of robotics and autonomous systems, developing accurate action models—internal representations of cause and effect—is crucial for effective planning and control. Traditional approaches often rely on the "smoothness assumption," which posits that numerically similar states lead to similar outcomes. While effective for continuous phenomena like gravity, these models fail in the presence of physical discontinuities, such as collisions. In these instances, small changes in state (e.g., a ball being millimeters away from a wall versus touching it) result in drastically different behaviors. This paper presents a hybrid learning system that integrates symbolic logic with continuous regression to model these complex, non-smooth environments.

Methodology: The Modal Model

The proposed system addresses discontinuities by identifying qualitatively distinct behaviors, termed modes. The learning process is divided into two primary components: segmentation and classification.

System Architecture

The overall architecture, illustrated in Figure 1, demonstrates the flow from environmental observation to prediction through two distinct phases: Learning and Prediction.

System architecture showing the learning and prediction phases.
  1. Segmentation (Finding the Modes): The system utilizes Expectation Maximization (EM) to cluster unlabeled training data into discrete linear functions. For instance, it might distinguish between the linear dynamics of a projectile in free-fall versus its dynamics during a collision. Forward stepwise linear regression is employed to maintain model simplicity and prevent overfitting.
  2. Classification (Selecting the Mode): Once modes are established, the agent must determine which mode to apply in a given state. This is achieved through a hybrid classifier:
  3. Symbolic: The FOIL (First-Order Inductive Learner) algorithm learns rules based on spatial predicates (e.g., intersect(ball, floor)).
  4. Numeric: Linear Discriminant Analysis (LDA) is used to resolve ambiguities that symbolic rules cannot capture, such as distinguishing between bouncing and rolling based on velocity.

Experimental Evaluation

The system was evaluated using a 2D physics simulation (Chipmunk Physics Engine) featuring a ball, a stationary box, and a ramp. To test generalization, the researchers randomized object positions, sizes, and initial velocities across 120 training and 120 test scenarios. Crucially, the coordinate system's origin was also randomized to ensure the model learned relative spatial relationships rather than absolute coordinates.

Key Results

The "Modal Model" (MM) was compared against Locally Weighted Regression (LWR), a standard baseline. The MM outperformed LWR by several orders of magnitude in velocity prediction. While LWR often struggled to exceed a null baseline, the MM achieved near-zero error once modes were correctly identified.

Classification Performance

A critical finding was the superiority of the symbolic classification approach. As shown in Figure 6, the FOIL-based classifier reached higher peak accuracy and learned significantly faster than standard numeric methods like Support Vector Machines (SVM) and Nearest Neighbors (NN).

Comparison of classification accuracy between FOIL, SVM, and NN across different scenarios.

This advantage is particularly pronounced in "hard" cases involving complex interactions like ramps and collisions, where symbolic rules provide a more robust framework for generalization than purely numeric boundaries.

Discussion and Insights

The core insight of this work is that physical behavior is fundamentally determined by relationships between objects rather than their absolute global coordinates. By using variablized logic (e.g., intersect(x, y)), the model can transfer knowledge learned from one object pair to another. Furthermore, modeling physics as a library of simple, local functions (piecewise) rather than a single global function proves to be highly efficient.

Limitations and Future Work

Despite its success, the current implementation faces challenges. The search for new modes and the execution of the FOIL algorithm are computationally intensive, limiting real-time application. Additionally, the model assumes linear dynamics within each mode, which may not suffice for non-linear phenomena like fluid dynamics. Future research will focus on automating the discovery of symbolic predicates directly from raw data, reducing the reliance on hard-coded spatial relationships.

Summary updated

BibTeX
@inproceedings{xu2013learning,
  author = {Joseph Zhen Ying Xu and John E. Laird},
  title = {{Learning Integrated Symbolic and Continuous Action Models for Continuous Domains}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2013},
  doi = {10.1609/aaai.v27i1.8676},
}
Figure