Addressing the Symbol Grounding Problem with Constraints in Neuro-Symbolic Planning (Short Paper)

Aymeric Barbin, Federico Cerutti, Alfonso Emilio GereviniIPS/RiCeRcA/SPIRIT@AI*IA 2022 · OpenReview

Introduction

In the field of artificial intelligence, a significant divide exists between sub-symbolic processing (neural networks handling raw data) and symbolic reasoning (classical planning). Neuro-symbolic planning aims to bridge this gap by learning to map sensory inputs, such as images, into discrete logical symbols that can be processed by classical planners. However, this mapping is fraught with the "Symbol Grounding Problem" (SGP), where the connection between raw data and stable symbols is inconsistent. This paper proposes a novel framework to address SGP by enforcing logical invariants during the neural network's training phase.

The Challenge of Symbol Grounding

State-of-the-art architectures like LatPlan utilize Variational Autoencoders (VAEs) to translate pixel-level data into categorical vector representations. These vectors serve as the "grounded" symbols for Planning Domain Description Language (PDDL) models. A critical issue arises when these learned representations lack stability: two slightly different images of the same physical state may result in different symbolic representations. This inconsistency breaks the "identity assumption" essential for symbolic logic—the principle that a state remains the same unless an action changes it. Consequently, the resulting state-space graph becomes disconnected, leading to "dead ends" during the planning search phase.

Proposed Methodology: Enforcing Domain Invariants

To mitigate representation instability, the authors suggest incorporating domain-specific logical invariants directly into the neural network's loss function. An invariant is a rule that must hold true across all valid states in a domain. For example, in a sliding tile puzzle, a single tile cannot occupy two different positions simultaneously. By penalizing the network when it generates representations that violate these rules, the model is forced to learn more consistent and logically sound mappings.

The proposed approach involves three key components:

1. Automated Invariant Discovery

Because manually defining invariants for every new domain is labor-intensive, the authors propose using automated tools. By applying systems like Fast Downward or DISCOPLAN to the PDDL generated by a preliminary LatPlan model, the system can identify "probable" invariants. It is important to note that these are statistical candidates rather than absolute truths, as they are derived from the neural network's potentially noisy initial output. Treating them as candidates allows the system to handle the inherent uncertainty of statistical inference.

2. Fuzzy Logic Loss Penalties

To integrate discrete logical constraints into a continuous, gradient-descent-based training environment, the authors employ t-norm fuzzy logic, a technique popularized by Logic Tensor Networks (LTN). This allows rigid logical formulas to be translated into differentiable functions. If a latent variable violates a constraint, the inverse of the constraint's truth value is added to the network's loss function as a penalty. This provides a gradient that guides the model toward logically consistent representations without requiring hard-coded symbolic rules.

3. Greedy Constraint Training

The authors introduce "Greedy Constraint Training" (Algorithm 1), an augmented training loop designed to filter candidate invariants. The loop acts as a form of hyperparameter search that iteratively:

  • Trains the model with a set of constraints.
  • Extracts new potential invariants from the resulting PDDL.
  • Evaluates these invariants based on their ability to reduce "State Variance" (a metric for representation stability).
  • Retains only those constraints that successfully improve the model's performance.

Discussion and Future Work

A primary insight of this work is the recognition that statistically learned PDDL is inherently noisy. By treating discovered invariants as hypotheses to be verified through a training loop, the system can theoretically self-correct. However, this iterative process introduces significant computational overhead, as it requires multiple retraining cycles of the LatPlan architecture.

While this short paper focuses on the theoretical framework and mathematical formulation, the authors identify empirical evaluation as the critical next step. Future research will involve implementing the GreedyConstraintTraining loop on standard benchmarks, such as the 8-puzzle, to quantify its impact on state variance and planning efficiency. The ultimate goal is to verify if this automated self-correction successfully discovers ground-truth invariants and improves the reliability of neuro-symbolic planners in practice.

Conclusion

By bridging the gap between rigid symbolic logic and flexible neural learning through fuzzy logic constraints, this research offers a promising path toward more robust neuro-symbolic agents. The ability to automatically discover and enforce domain invariants could significantly reduce the manual engineering required to solve the Symbol Grounding Problem in complex, sub-symbolic environments.

Summary updated

BibTeX
@inproceedings{barbin2022addressing,
  author = {Aymeric Barbin and Federico Cerutti and Alfonso Emilio Gerevini},
  title = {{Addressing the Symbol Grounding Problem with Constraints in Neuro-Symbolic Planning (Short Paper)}},
  booktitle = {IPS/RiCeRcA/SPIRIT@AI*IA},
  year = {2022},
  url = {https://openreview.net/forum?id=iBlSsVQNLDa},
}
Figure