Cross-Entropy Admissibility (CEA)

• Cross-Entropy Admissibility (CEA) is a modified loss function that enforces admissibility in heuristic learning by penalizing overestimated cost-to-go values.
• CEA adjusts the standard cross-entropy loss by weighting only admissible classes, thus ensuring the heuristic remains informative while not overestimating true costs.
• Empirical results, particularly in Rubik’s Cube PDB domains, show that CEA-trained models achieve near-perfect admissibility and significantly lower overestimation rates compared to traditional methods.

Cross-Entropy Admissibility (CEA) most commonly denotes the loss function introduced for learning admissible heuristics for A*, where admissibility is the property of never overestimating the true shortest-path cost and therefore guarantees solution optimality. In that formulation, heuristic learning is cast as classification over cost-to-go classes, and the loss assigns credit only to classes at or below the true cost-to-go, with a small auxiliary cross-entropy term used for sharpening (Futuhi et al., 26 Sep 2025). This suggests a broader organizing idea: across several literatures, “admissibility” attached to cross-entropy refers to whether a cross-entropy-based objective preserves a domain-specific guarantee, such as A* optimality, likelihood equivalence, or certified surrogate validity (Shangnan et al., 2021, Héas, 2018).

1. Search-theoretic setting

In the A* setting, heuristic functions are central to performance, and admissibility means that the heuristic $h(\cdot)$ never overestimates the true cost-to-go $h^*(\cdot)$. Traditional heuristics such as Pattern Databases (PDBs) guarantee admissibility via domain abstraction, but are limited by storage constraints and sometimes lack informativeness. Recent deep learning approaches often disregard admissibility and provide limited guarantees on generalization beyond the training data (Futuhi et al., 26 Sep 2025).

The central asymmetry is that overestimation endangers optimality, whereas underestimation only affects efficiency. Standard losses such as mean squared error or regular cross-entropy treat over- and under-estimation symmetrically, so they do not directly encode the operational requirement imposed by A*. CEA was introduced precisely to address this mismatch by posing heuristic learning as a constrained optimization problem and enforcing admissibility during training. In the Rubik’s Cube domain, the resulting heuristics were reported as near-admissible and as providing significantly stronger guidance than compressed PDB heuristics (Futuhi et al., 26 Sep 2025).

2. Formal definition of the CEA loss

When heuristic learning is treated as classification, the true cost-to-go $h^*(s)$ becomes a class label. The standard cross-entropy loss is

$$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$

where $p_k^{(i)}$ is the predicted probability for class $k$ on example $i$, and $y_k^{(i)}$ is 1 if the true class for $i$ is $k$, else 0. Its limitation is structural: it treats over- and under-estimations equally (Futuhi et al., 26 Sep 2025).

CEA modifies this objective to bias the model toward admissible predictions while maintaining classification sharpness:

$h^*(\cdot)$0

Here $h^*(\cdot)$1 is the number of samples, $h^*(\cdot)$2 is the true cost-to-go for sample $h^*(\cdot)$3, $h^*(\cdot)$4 is the predicted probability for class $h^*(\cdot)$5, $h^*(\cdot)$6 controls the trade-off between mean heuristic value and admissibility, and $h^*(\cdot)$7 is a regularization weight for the standard CE term. The first term sums prediction probabilities only over classes $h^*(\cdot)$8, and weights them by $h^*(\cdot)$9, thereby preferring higher but still admissible heuristic values. The second term encourages sharpness at the correct class without dominating the admissibility bias (Futuhi et al., 26 Sep 2025).

3. Admissibility mechanism and optimization behavior

CEA enforces admissibility by withholding credit from any probability mass assigned to an inadmissible class $h^*(s)$0. In the main term of the loss, only admissible classes contribute, so minimization favors distributions supported on $h^*(s)$1. The weighting $h^*(s)$2 further encourages the highest possible admissible class, making the heuristic as informative as possible while respecting the A* constraint (Futuhi et al., 26 Sep 2025).

At the global optimum, $h^*(s)$3 for all $h^*(s)$4, so the model is maximally informative and perfectly admissible. The auxiliary CE penalty with weight $h^*(s)$5 prevents the model from distributing mass too freely across all admissible classes. This design distinguishes CEA from post hoc clipping or masking strategies, because admissibility is encoded directly in the training objective rather than imposed only at inference time. The paper characterizes CEA as the first practical, scalable neural loss function to directly penalize and minimize inadmissibility (Futuhi et al., 26 Sep 2025).

4. Sample complexity and generalization

The theoretical analysis couples admissible heuristic learning to structural properties of the search graph and to the capacity of the hypothesis class. By leveraging PDB abstractions, the sample complexity can depend on the size of the compressed state space rather than the size of the full combinatorial graph. If the PDB abstraction reduces the state space from $h^*(s)$6 to $h^*(s)$7, the sample complexity is $h^*(s)$8, which the paper describes as a potentially exponential reduction (Futuhi et al., 26 Sep 2025).

For a ReLU neural network of size $h^*(s)$9, depth $$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$0, width $$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$1, and $$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$2 heuristic classes, the pseudo-dimension is

$$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$3

Replacing a general hypothesis class with a ReLU neural network therefore gives bounds that depend primarily on network width and depth, rather than on graph size. The analysis also provides generalization guarantees for goal-dependent heuristics, stated as the first such guarantees in the paper (Futuhi et al., 26 Sep 2025).

The key performance functional is the maximum inadmissibility encountered on optimal paths. For this quantity, the expected test-time inadmissibility is bounded by the empirical training inadmissibility plus a complexity-controlled term. The practical upshot reported in the paper is that when training overestimation is near zero, test-time inadmissibility is also tightly controlled, so CEA-trained models are provably nearly-admissible on unseen instances if enough data is provided (Futuhi et al., 26 Sep 2025).

5. Empirical behavior on Rubik’s Cube PDB domains

CEA was evaluated on several Rubik’s Cube PDB domains: 8-corner, 7-edge, 6-edge, and delta edge PDBs. The baselines were compressed PDBs using admissible min compression under the same memory budget as the neural heuristics, and standard cross-entropy neural heuristics. The reported result is that CEA consistently outperformed standard CE by several orders of magnitude in overestimation rate, and that CEA-trained heuristics were stronger than same-sized compressed PDBs (Futuhi et al., 26 Sep 2025).

For the 8-corner PDB, the paper reports the following comparison.

Heuristic	Avg. Heuristic	Overestimation Rate
NN + CEA loss	8.76	$$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$4
NN + CE loss	8.76	$$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$5
Compressed PDB	6.84	0

The same 8-corner experiment used model sizes of 1.89 MB for the two neural models and 1.91 MB for the compressed PDB, with compression rates of 23.3×, 23.3×, and 23.0× respectively. For the 6-edge and 7-edge PDBs, overestimation rates with CEA were reported as $$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$6–$$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$7. In some cases, notably 8-corner PDB, CEA learned a perfectly admissible heuristic that matched the original, uncompressed PDB. Across the reported experiments, neural models trained with CEA achieved up to 51× memory compression relative to the original PDB while retaining nearly perfect admissibility and informativeness. Performance was also described as stable across a range of $$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$8 and $$\text{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{k=1}^{\ell} y_k^{(i)}\log p_k^{(i)},$$9 values when using a staged training schedule (Futuhi et al., 26 Sep 2025).

6. Other task-specific admissibility notions around cross-entropy

In quantum machine learning, admissibility is tied to whether minimizing quantum cross-entropy coincides with maximizing likelihood. The quantum cross-entropy is defined as

$p_k^{(i)}$0

with $p_k^{(i)}$1 when the support condition fails. In the undisturbed regime, where the learning algorithm can access the true quantum state directly, minimizing quantum cross-entropy is equivalent to maximizing quantum likelihood:

$p_k^{(i)}$2

When quantum cross-entropy is constructed from empirical density matrices assembled from projective measurement outcomes, the relation weakens to a lower bound,

$p_k^{(i)}$3

reflecting information loss under measurement. Equality holds if and only if $p_k^{(i)}$4 commutes with all measurement projectors and the measurement projectors restricted to the support of $p_k^{(i)}$5 are rank-1. The paper concludes that the deferred measurement principle is crucial for “full quantum machine learning” because it preserves the better-behaved, undisturbed quantum cross-entropy (Shangnan et al., 2021).

In rare-event simulation, admissibility appears in a different sense: a reduced model is admissible at a given CE iteration if its certified approximation error does not compromise the quantile gap needed for nested-event progression. With high-fidelity score function $p_k^{(i)}$6, surrogate $p_k^{(i)}$7, and local certificate $p_k^{(i)}$8, the paper uses

$p_k^{(i)}$9

to select the lowest-dimensional surrogate whose accuracy is sufficient for the current CE stage. Under the stated assumptions, the resulting algorithm converges almost surely to the optimal biasing distribution in at most as many iterations as the standard CE method, and at each stage achieves the minimal possible variance for its chosen level of surrogate approximation (Héas, 2018).

7. Scope, ambiguities, and open issues

The available literature suggests that “cross-entropy admissibility” is not a single universal definition but a family of domain-dependent criteria. In A* heuristic learning, admissibility means never overestimating $k$0. In quantum learning, it means preserving the equivalence between cross-entropy minimization and likelihood maximization. In reduced-model CE, it means certifying that a surrogate preserves the optimization stage. This suggests that CEA is best understood through the operational guarantee it is meant to preserve, rather than through a single abstract definition.

Related information-theoretic discussions reinforce that admissibility can depend on the exact cross-entropy variant and the decision-theoretic setting. For the Rényi cross-entropy, the summarized analysis states that $k$1 is minimized at $k$2, but that strict convexity and strict properness may fail for some values of $k$3; it also notes that the differential Rényi cross-entropy can be negative and possibly unbounded, so the continuous case requires care (Thierrin et al., 2022). By contrast, the streaming generalized cross-entropy literature explicitly does not provide a formal theorem or definition regarding admissibility of the estimator; it instead discusses the balance between prior knowledge and new data, structural bias in streaming updates, and open questions concerning optimal weights and update rules (Angelelli et al., 2018).

A further terminological caution arises from neighboring work on entropy admissibility in nonlocal traffic models, where the issue is not cross-entropy as a loss but the selection of the entropy admissible solution in a singular local limit. There, strong $k$4 convergence of $k$5 to $k$6 implies that the limit is the entropy admissible solution, while convexity of the kernel is crucial for the compactness estimate that yields that convergence (Colombo et al., 2022). The conceptual overlap is therefore limited: the shared term is “admissibility,” but the mathematical role of entropy is different.

Within the A* literature itself, a common misconception is that ordinary cross-entropy is already adequate once heuristic prediction is posed as classification. The CEA construction was introduced precisely because ordinary CE does not penalize inadmissibility asymmetrically. Another misconception is that near-admissibility and perfect admissibility are interchangeable. The reported empirical record includes both near-admissible heuristics and, in some cases, perfectly admissible heuristics, whereas the theoretical analysis emphasizes sample complexity and generalization control of inadmissibility rather than an unconditional guarantee of perfection on unseen states (Futuhi et al., 26 Sep 2025).