arXiv subjects

Dimitris Fouskakis

Publications and source records attributed to Dimitris Fouskakis.

9 recordsLinked to original sources

arXiv2026

Position: agentic AI orchestration should be Bayes-consistent

LLMs excel at predictive tasks and complex reasoning tasks, but many high-value deployments rely on decisions under uncertainty, for example, which tool to call, which expert to consult, or how many resources to invest. While the usefulness and feasibility of Bayesian approaches remain unclear for LLM inference, this position paper argues that the control layer of an agentic AI system (that orchestrates LLMs and tools) is a clear case where Bayesian principles should shine. Bayesian decision theory provides a framework for agentic systems that can help to maintain beliefs over task-relevant latent quantities, to update these beliefs from observed agentic and human-AI interactions, and to choose actions. Making LLMs themselves explicitly Bayesian belief-updating engines remains computationally intensive and conceptually nontrivial as a general modeling target. In contrast, this paper argues that coherent decision-making requires Bayesian principles at the orchestration level of the agentic system, not necessarily the LLM agent parameters. This paper articulates practical properties for Bayesian control that fit modern agentic AI systems and human-AI collaboration, and provides concrete examples and design patterns to illustrate how calibrated beliefs and utility-aware policies can improve agentic AI orchestration.

cs.AI↗

arXiv2020

Power-Expected-Posterior Priors as Mixtures of g-Priors

One of the main approaches used to construct prior distributions for objective Bayes methods is the concept of random imaginary observations. Under this setup, the expected-posterior prior (EPP) offers several advantages, among which it has a nice and simple interpretation and provides an effective way to establish compatibility of priors among models. In this paper, we study the power-expected posterior prior as a generalization to the EPP in objective Bayesian model selection under normal linear models. We prove that it can be represented as a mixture of $g$-prior, like a wide range of prior distributions under normal linear models, and thus posterior distributions and Bayes factors are derived in closed form, keeping therefore computational tractability. Comparisons with other mixtures of $g$-prior are made and emphasis is given in the posterior distribution of g and its effect on Bayesian model selection and model averaging.

stat.ME↗

arXiv2016

Variations of Power-Expected-Posterior Priors in Normal Regression Models

The power-expected-posterior (PEP) prior is an objective prior for Gaussian linear models, which leads to consistent model selection inference, under the M-closed scenario, and tends to favor parsimonious models. Recently, two new forms of the PEP prior were proposed which generalize its applicability to a wider range of models. The properties of these two PEP variants within the context of the normal linear model are examined thoroughly, focusing on the prior dispersion and on the consistency of the induced model selection procedure. Results show that both PEP variants have larger variances than the unit-information g-prior and that they are M-closed consistent as the limiting behavior of the corresponding marginal likelihoods matches that of the BIC. The consistency under the M-open case, using three different model misspecification scenarios is further investigated.

stat.ME↗

arXiv2015

Power-Expected-Posterior Priors for Generalized Linear Models

The power-expected-posterior (PEP) prior provides an objective, automatic, consistent and parsimonious model selection procedure. At the same time it resolves the conceptual and computational problems due to the use of imaginary data. Namely, (i) it dispenses with the need to select and average across all possible minimal imaginary samples, and (ii) it diminishes the effect that the imaginary data have upon the posterior distribution. These attributes allow for large sample approximations, when needed, in order to reduce the computational burden under more complex models. In this work we generalize the applicability of the PEP methodology, focusing on the framework of generalized linear models (GLMs), by introducing two new PEP definitions which are in effect applicable to any general model setting. Hyper-prior extensions for the power parameter that regulates the contribution of the imaginary data are introduced. We further study the validity of the predictive matching and of the model selection consistency, providing analytical proofs for the former and empirical evidence supporting the latter. For estimation of posterior model and inclusion probabilities we introduce a tuning-free Gibbs-based variable selection sampler. Several simulation scenarios and one real life example are considered in order to evaluate the performance of the proposed methods compared to other commonly used approaches based on mixtures of g-priors. Results indicate that the GLM-PEP priors are more effective in the identification of sparse and parsimonious model formulations.

stat.ME↗

arXiv2014

A Bayesian Hierarchical Model for Comparative Evaluation of Teaching Quality Indicators in Higher Education

The problem motivating the paper is the quantification of students' preferences regarding teaching/coursework quality, under certain numerical restrictions, in order to build a model for identifying, assessing and monitoring the major components of the overall academic quality. After reviewing the strengths and limitations of conjoint analysis and of the random coefficient regression model used in similar problems in the past, we propose a Bayesian beta regression model with a Dirichlet prior on the model coefficients. This approach not only allows for the incorporation of informative prior when it is available but also provides user friendly interfaces and direct probability interpretations for all quantities. Furthermore, it is a natural way to implement the usual constraints for the model weights/coefficients. This model was applied to data collected in 2009 and 2013 from undergraduate students in Panteion University, Athens, Greece and besides the construction of an instrument for the assessment and monitoring of teaching quality, it gave some input for a preliminary discussion on the association of the differences in students preferences between the two time periods with the current Greek economic and financial crisis.

stat.AP↗

arXiv2013

Bayesian transformation family selection: moving towards a transformed Gaussian universe

The problem of transformation selection is thoroughly treated from a Bayesian perspective. Several families of transformations are considered with a view to achieving normality: the Box-Cox, the Modulus, the Yeo & Johnson and the Dual transformation. Markov chain Monte Carlo algorithms have been constructed in order to sample from the posterior distribution of the transformation parameter $\lambda_T$ associated with each competing family $T$. We investigate different approaches to constructing compatible prior distributions for $\lambda_T$ over alternative transformation families, using a unit-information power-prior approach and an alternative normal prior with approximate unit-information interpretation. Selection and discrimination between different transformation families is attained via posterior model probabilities. We demonstrate the efficiency of our approach using a variety of simulated datasets. Although there is no choice of transformation family that can be universally applied to all problems, empirical evidence suggests that some particular data structures are best treated by specific transformation families. For example, skewness is associated with the Box-Cox family while fat-tailed distributions are efficiently treated using the Modulus transformation.

stat.ME↗

arXiv2013

Limiting behavior of the Jeffreys Power-Expected-Posterior Bayes Factor in Gaussian Linear Models

Expected-posterior priors (EPP) have been proved to be extremely useful for testing hypothesis on the regression coefficients of normal linear models. One of the advantages of using EPPs is that impropriety of baseline priors causes no indeterminacy. However, in regression problems, they based on one or more \textit{training samples}, that could influence the resulting posterior distribution. The power-expected-posterior priors are minimally-informative priors that diminishing the effect of training samples on the EPP approach, by combining ideas from the power-prior and unit-information-prior methodologies. In this paper we show the consistency of the Bayes factors when using the power-expected-posterior priors, with the independence Jeffreys (or reference) prior as a baseline, for normal linear models under very mild conditions on the design matrix.

stat.CO↗

arXiv2013

Power-Conditional-Expected Priors: Using g-priors with Random Imaginary Data for Variable Selection

The Zellner's g-prior and its recent hierarchical extensions are the most popular default prior choices in the Bayesian variable selection context. These prior set-ups can be expressed power-priors with fixed set of imaginary data. In this paper, we borrow ideas from the power-expected-posterior (PEP) priors in order to introduce, under the g-prior approach, an extra hierarchical level that accounts for the imaginary data uncertainty. For normal regression variable selection problems, the resulting power-conditional-expected-posterior (PCEP) prior is a conjugate normal-inverse gamma prior which provides a consistent variable selection procedure and gives support to more parsimonious models than the ones supported using the g-prior and the hyper-g prior for finite samples. Detailed illustrations and comparisons of the variable selection procedures using the proposed method, the g-prior and the hyper-g prior are provided using both simulated and real data examples.

stat.CO↗

arXiv2013

Power-Expected-Posterior Priors for Variable Selection in Gaussian Linear Models

In the context of the expected-posterior prior (EPP) approach to Bayesian variable selection in linear models, we combine ideas from power-prior and unit-information-prior methodologies to simultaneously produce a minimally-informative prior and diminish the effect of training samples. The result is that in practice our power-expected-posterior (PEP) methodology is sufficiently insensitive to the size n* of the training sample, due to PEP's unit-information construction, that one may take n* equal to the full-data sample size n and dispense with training samples altogether. In this paper we focus on Gaussian linear models and develop our method under two different baseline prior choices: the independence Jeffreys (or reference) prior, yielding the J-PEP posterior, and the Zellner g-prior, leading to Z-PEP. We find that, under the reference baseline prior, the asymptotics of PEP Bayes factors are equivalent to those of Schwartz's BIC criterion, ensuring consistency of the PEP approach to model selection. We compare the performance of our method, in simulation studies and a real example involving prediction of air-pollutant concentrations from meteorological covariates, with that of a variety of previously-defined variants on Bayes factors for objective variable selection. Our prior, due to its unit-information structure, leads to a variable-selection procedure that (1) is systematically more parsimonious than the basic EPP with minimal training sample, while sacrificing no desirable performance characteristics to achieve this parsimony; (2) is robust to the size of the training sample, thus enjoying the advantages described above arising from the avoidance of training samples altogether; and (3) identifies maximum-a-posteriori models that achieve good out-of-sample predictive performance.

stat.CO↗