Bayesian Symbolic Agent (EMPA)

• Bayesian Symbolic Agent (EMPA) is a computational framework that integrates symbolic reasoning with Bayesian inference by modeling logical formulas as random variables influenced by empirical data.
• The architecture leverages a tunable abstraction parameter to interpolate between classical, empirical, and paraconsistent reasoning, enhancing flexibility across various applications.
• EMPA employs scalable Bayesian updates and active exploration, with neuro-symbolic variants demonstrating robust performance in cognitive modeling and robotics.

A Bayesian Symbolic Agent (EMPA) is a computational framework unifying symbolic reasoning and Bayesian probabilistic inference. EMPA treats logical formulas as random variables whose truth depends on latent, data-driven models. By tuning a single abstraction parameter, EMPA interpolates between classical logic, empirical generalization, paraconsistent reasoning, and fully probabilistic forms of inference. This architecture provides a generic foundation for constructing agents capable of learning and robust reasoning over symbolic structures derived from empirical data, applicable across domains from cognitive modeling to robotics (Kido, 2024, Arriaga et al., 10 Jun 2025, Andersen et al., 2017).

1. Theoretical Foundations and Formalism

The conceptual core of EMPA is a single probabilistic abstraction principle: symbolic knowledge arises by treating the truth of logical formulas as random variables determined by latent states (models) supported by observed data (Kido, 2024). The basic setting consists of:

• A finite propositional language $L$ whose atoms admit $N$ complete truth-value assignments (models $m_1, ..., m_N$).
• A corpus of data $D = \{d_1 , ..., d_K\}$, where each datum $d_k$ uniquely supports a model $m(d_k)$.
• Logical formulas $\alpha \in L$ are associated with random variables whose value is $1$ (true) or $0$ (false), depending on the sampled model.

EMPA defines several consequence relations, all unified under its probabilistic abstraction:

• Classical consequence: $\Delta \models \alpha$ if every model satisfying $N$0 satisfies $N$1.
• Empirical consequence: $N$2 if all possible models of $N$3 (i.e., those with $N$4) are also possible models of $N$5.
• Paraconsistent and parapossible reasoning: When $N$6 is inconsistent or impossible, inference is performed over all maximal consistent (MCS) or maximal possible subsets (MPS) of $N$7 respectively, to recover robust reasoning in the presence of contradiction or impossibility.

2. Mathematical and Algorithmic Structure

The generative model underpinning EMPA introduces:

• A hidden world-state random variable $N$8 with empirical likelihood $N$9, where $m_1, ..., m_N$0 is the number of data supporting $m_1, ..., m_N$1.
• Formula likelihood: for each formula $m_1, ..., m_N$2, $m_1, ..., m_N$3, where $m_1, ..., m_N$4 is $m_1, ..., m_N$5 if $m_1, ..., m_N$6, $m_1, ..., m_N$7 otherwise.
• Marginals and conditionals: $m_1, ..., m_N$8; $m_1, ..., m_N$9.

The key abstraction parameter $D = \{d_1 , ..., d_K\}$0 offers a spectrum:

• $D = \{d_1 , ..., d_K\}$1 recovers deterministic logic;
• $D = \{d_1 , ..., d_K\}$2 yields statistical inference.

Bayesian update is achieved via $D = \{d_1 , ..., d_K\}$3 on observing new data (Kido, 2024).

The high-level loop comprises:

1. Data acquisition (observe $D = \{d_1 , ..., d_K\}$4).
2. Model update ($D = \{d_1 , ..., d_K\}$5; $D = \{d_1 , ..., d_K\}$6).
3. Formula update (compute $D = \{d_1 , ..., d_K\}$7).
4. Query answering (compute $D = \{d_1 , ..., d_K\}$8, applying special-casing for paraconsistency/parapossibility).

3. Reasoning Modes and Unification

By tuning $D = \{d_1 , ..., d_K\}$9 and considering different consistency regimes, EMPA integrates several reasoning paradigms:

• Classical entailment via supported and consistent premises ($d_k$0).
• Statistical inference when $d_k$1.
• Paraconsistent reasoning (contradictory premises, $d_k$2) via MCS.
• Parapossible reasoning (impossible premises, $d_k$3) via MPS.

This flexibility allows graded, robust, and data-driven symbolic inference, with inference over largest coherent subsets providing resilience to contradiction and incompleteness. Selective attention and abstraction via maximal subsets is analogous to cognitive strategies observed in humans (Kido, 2024).

4. EMPA Variants and Symbolic Model Learning

Multiple instantiations of EMPA have been proposed:

A. Active Exploration for Symbolic Representation Learning

EMPA supports online construction and Bayesian refinement of symbolic state spaces and option models. The agent maintains a posterior $d_k$4 over symbolic models; each model $d_k$5 specifies clustered symbolic state factors and action precondition/effect distributions. Bayesian parameter updates employ Beta and sparse Dirichlet priors, with posterior factorization across options and symbolic state partitions.

Active exploration is guided by maximizing expected information gain (reduction in posterior standard deviation $d_k$6), implemented via Monte-Carlo Tree Search (MCTS) for tractable policy optimization (Andersen et al., 2017).

Empirical evaluation in benchmark domains (Asteroids, Treasure Game) demonstrates data efficiency and coverage that significantly surpass random or greedy exploration baselines, requiring about one-third as many action executions to achieve equivalent symbolic model completeness.

B. Neuro-symbolic Hybrid EMPA

In robotic learning contexts, EMPA is instantiated as a modular agent unifying:

• Differentiable-physics world model $d_k$7,
• Bayesian inference over physical parameters $d_k$8,
• Meta-learning for rapid task adaptation (MAML-style optimization over task batches),
• Symbolic program synthesis layer for high-level action composition.

Probabilistic evaluation of symbolic preconditions and postconditions is grounded in current posteriors over continuous latent states, and symbolic planning is adapted to explicit model uncertainty (e.g., via RJMCMC-based synthesis). The integrated planning/inference loop propagates Bayesian posteriors as signals for uncertainty-aware high-level control (Arriaga et al., 10 Jun 2025).

5. Scalability and Computational Features

Algorithmic efficiency in EMPA is achieved by structuring probabilistic inference around data-supported models rather than all possible truth assignments, reducing the computational burden from $d_k$9 to $m(d_k)$0, where $m(d_k)$1 is the number of observed data points and $m(d_k)$2 the formula set size (Kido, 2024). In symbolic model construction, the factorization of posteriors and the use of conjugate updates enable closed-form Bayesian refinement, scalable to online learning scenarios (Andersen et al., 2017).

Meta-learning variants leverage batch optimization across tasks and amortized inference, facilitating rapid adaptation without excessive retraining (Arriaga et al., 10 Jun 2025).

6. Extensions and Practical Implications

EMPA accommodates several extensions:

• Lifted inference over first-order logics via tractable model-counting.
• Continuous latent-state handling through approximate inference techniques (e.g., variational Bayes, HMC).
• Integration with utility-theoretic decision-making, facilitating probabilistic symbolic planning and active data acquisition.
• Domain-specific leveraging of symbolic abstraction for efficient robotic world modeling and adaptive planning (Kido, 2024, Arriaga et al., 10 Jun 2025).

A plausible implication is that, by formalizing symbolic abstraction as a statistical process over data-dependent latent states, EMPA provides a unified substrate for constructing agents with flexible, scalable, and robust symbolic reasoning, directly linking empirical observations to high-level logical operations.

7. Empirical Results and Domain Applications

EMPA’s empirical performance has been quantitatively validated in symbolic exploration tasks and conceptualized for use in robotics:

• In game domains, EMPA achieves faster coverage of symbolic transition spaces and more efficient exploration than undirected baselines, as measured by fraction of unobserved symbolic transitions and successful attainment of rare object states (Andersen et al., 2017).
• In neuro-symbolic robotics, EMPA-based architectures are proposed to enable uncertainty-aware symbolic planning and adaptive generalization, integrating learned physical models, meta-learned residuals, and probabilistic symbolic program synthesis for interpretable and data-efficient control in novel environments (Arriaga et al., 10 Jun 2025).

These results support EMPA’s suitability as a principled and extensible agent framework for diverse symbolic reasoning and learning tasks.