ANTR: Posterior Transformation with Adaptive Trust-Regions

• The paper demonstrates that ANTR reduces parameter estimation error by up to 50% while using only 36–53% of the simulation budget compared to baselines.
• It employs a neural density estimator for amortized Bayesian inference combined with negatively correlated search to maintain search diversity in complex, multimodal spaces.
• Adaptive trust-regions dynamically resize based on local success rates, optimizing resource allocation for efficient calibration of agent-based models.

Automatic Posterior Transformation with Negatively Correlated Search and Adaptive Trust-Region (ANTR) is a surrogate-based optimization framework for efficiently calibrating agent-based models (ABMs), particularly in applications where evaluation is expensive and the parameter space is nonlinear and multimodal. The method leverages a neural density estimator to directly approximate the Bayesian posterior over parameters, implements a negatively correlated search to enhance diversity among search agents, and applies adaptive trust-region strategies for efficient local search and resource allocation. The framework is specifically designed to address the challenges of batched calibration tasks across heterogeneous agent-based financial market simulators (Jiang et al., 11 Jan 2026).

1. Mathematical Foundations

ANTR formalizes the calibration problem for ABMs as the identification of parameters $\theta \in \Theta \subseteq \mathbb{R}^d$ that minimize the discrepancy $D(M(\theta), x_{\mathrm{obs}})$ between simulator output $M(\theta)$ and observed data $x_{\mathrm{obs}}$, where $M(\theta)$ is typically a time series. In Bayesian terms, the posterior $p(\theta|x_{\mathrm{obs}})$ involves an intractable likelihood $p(x_{\mathrm{obs}}|\theta)$. ANTR circumvents this by training a neural surrogate $q_\phi(\theta|x)$ to approximate the posterior via simulated $(\theta, x)$ pairs.

Surrogate training minimizes KL divergence to the true joint via

$$\mathcal{L}(\phi) = \mathbb{E}_{p(\theta, x)}[-\log q_\phi(\theta | x)] \approx -\frac{1}{N} \sum_{i=1}^{N} \log q_\phi(\theta_i | x_i)$$

where data is generated by sampling parameters from a prior and propagating through the simulator.

2. Neural Posterior Modeling and Amortization

Central to ANTR is the Automatic Posterior Transformation (APT) surrogate, which consists of two primary modules: a sequence embedder $z = f_{\mathrm{enc}}(x)$—employing a CNN for time series—and a conditional density estimator $q_\phi(\theta|z)$, instantiated as either a Mixture Density Network or a Normalizing Flow.

The surrogate is pretrained with a large amortization dataset (e.g., 19,200 samples for the Brock–Hommes model, 120,000 for the Preis–Golke–Paul–Schneider model). Training is conducted via stochastic gradient descent over mini-batches. Once trained, $q_\phi$ supports amortized inference, allowing rapid approximation of posteriors for distinct observed datasets from the same ABM with minimal retraining.

3. Negatively Correlated Search Dynamics

Within each trust region, ANTR maintains $N$ Randomized Local Search (RLS) agents, each defined by a Gaussian search distribution $p_i(\theta) = \mathcal{N}(\theta_i, \Sigma_i)$. Each agent proposes candidate parameters, evaluates their surrogate log-posterior $$f(\theta_i') = \log q_\phi(\theta_i' | x_{\mathrm{obs}})$$, and quantifies diversity via the Bhattacharyya distance to other agents' distributions.

A diversity-driven replacement rule, governed by hyperparameter $\lambda$,

$$\text{if } \frac{f(\theta_i')}{\mathrm{Corr}(p_i')} < \lambda, \text{ then update} (\theta_i, \Sigma_i) \leftarrow (\theta_i', \Sigma_i)$$

prevents collapse to a single mode and enforces negatively correlated exploration.

4. Adaptive Trust-Region Mechanism

ANTR partitions the parameter space into $M$ hyperrectangular trust regions $\{\mathcal{T}_j\}_{j=1}^M$, each with dedicated surrogate retraining and search population. The success ($c_{s,j}$) and failure ($c_{f,j}$) counters within each region modulate expansion (doubling) and contraction (halving) of region sizes: $$\Delta_j \leftarrow \begin{cases} 2\Delta_j & \text{if } c_{s,j}\ge3 \ \Delta_j/2 & \text{if } c_{f,j}\geq \mathrm{fail\_tol} \end{cases}$$ with $\mathrm{fail\_tol} = \lceil \max(4/N, d/N) \rceil$. Regions can be retired or respawned based on calibration progress.

5. Algorithmic Workflow

ANTR's full workflow entails the following stages (summarized below in table format):

Stage	Core Procedure	Resource
Initialization	Pretrain $q_\phi$. Identify $M$ high-density seeds.	One-off dataset
Trust Regions	Parallel search and local surrogate updates.	$M$ regions
NCS Iterations	$N$ searchers per region, diversity preserving.	$M\times N$ RLS
Simulator Calls	Evaluate candidates $M\times N$ per iteration.	Expensive batch
Adaptation	Trust-region resizing and possible respawning.	Dynamic

Each iteration collects new simulation data, refines local surrogates, executes NCS steps, and adapts trust regions based on success/failure counters. Convergence criteria are based on iteration budget or region size thresholds.

6. Empirical Evaluation

ANTR's performance is assessed on two agent-based financial market models:

• Brock–Hommes (BH) Model: Calibration in 2D and 4D, budgeted at 900 and 1800 simulator evaluations, respectively, with $T=900$ and $T=1800$. Time series length $T=900$.
• Preis–Golke–Paul–Schneider (PGPS) Model: Six-dimensional parameter space, data lengths 600–3600.

Metrics include mean squared error (MSE), Euclidean parameter error, success rate (fraction within $d \le r\sqrt{d}$), and relative evaluation budget.

Key findings:

• For 10 BH test cases, ANTR achieved the lowest MSE in 6 cases and smallest parameter error in 8 cases; win–tie–loss vs. TuRBO: 6–0–4 (MSE), 8–0–2 (parameter error).
• ANTR's average parameter error reduced by approximately 50% under tight budgets.
• To match or surpass baselines, ANTR achieved comparable accuracy with only 36–53% of the simulation budget.
• On PGPS, ANTR reached up to 100% success rate and consistently smaller parameter error, outperforming baselines operating between 0–90% success.

7. Limitations, Methodological Insights, and Application Guidelines

The ANTR framework deploys heuristic strategies in trust-region management and lacks explicit mechanisms to target high surrogate–true discrepancy for active learning. Potential extensions include integration of acquisition-function-based active learning for NCS and meta-adaptation for core hyperparameters ($\lambda$, $\mathrm{fail\_tol}$, trust-region scaling). Online refinement of $q_\phi$ can further improve adaptation when new observed data becomes available.

For effective application:

• Curate sufficiently large amortization datasets for the ABM under study to ensure robust global posterior estimation.
• Choose embedding architectures suited to the data modality (CNNs for time series, GNNs for graphs, etc.).
• Carefully tune NCS parameters ($N$, $\lambda$) to preserve diversity, especially in high-dimensional settings.
• Configure $M \approx \lceil \frac{\text{budget}}{100N} \rceil$ to balance regional specialization and resource fragmentation.
• Secure adequate parallel resources to execute $M\times N$ simulator evaluations per iteration.

ANTR exhibits state-of-the-art calibration accuracy and sample efficiency across nonlinear, multimodal ABM calibration tasks by uniting amortized neural posterior estimation, negatively correlated evolutionary search, and adaptive trust-regions (Jiang et al., 11 Jan 2026).