Albert Thomas

Giuseppe Paolo, Jonas Gonzalez-Billandon, Albert Thomas et al.

In the last decade, reinforcement learning successfully solved complex control tasks and decision-making problems, like the Go board game. Yet, there are few success stories when it comes to deploying those algorithms to real-world scenarios. One of the reasons is the lack of guarantees when dealing with and avoiding unsafe states, a fundamental requirement in critical control engineering systems. In this paper, we introduce Guided Safe Shooting (GuSS), a model-based RL approach that can learn to control systems with minimal violations of the safety constraints. The model is learned on the data collected during the operation of the system in an iterated batch fashion, and is then used to plan for the best action to perform at each time step. We propose three different safe planners, one based on a simple random shooting strategy and two based on MAP-Elites, a more advanced divergent-search algorithm. Experiments show that these planners help the learning agent avoid unsafe situations while maximally exploring the state space, a necessary aspect when learning an accurate model of the system. Furthermore, compared to model-free approaches, learning a model allows GuSS reducing the number of interactions with the real-system while still reaching high rewards, a fundamental requirement when handling engineering systems.

Geovani Rizk, Igor Colin, Albert Thomas et al.

We propose the first regret-based approach to the Graphical Bilinear Bandits problem, where $n$ agents in a graph play a stochastic bilinear bandit game with each of their neighbors. This setting reveals a combinatorial NP-hard problem that prevents the use of any existing regret-based algorithm in the (bi-)linear bandit literature. In this paper, we fill this gap and present the first regret-based algorithm for graphical bilinear bandits using the principle of optimism in the face of uncertainty. Theoretical analysis of this new method yields an upper bound of $\tilde{O}(\sqrt{T})$ on the $α$-regret and evidences the impact of the graph structure on the rate of convergence. Finally, we show through various experiments the validity of our approach.

Abdelhakim Benechehab, Vasilii Feofanov, Giuseppe Paolo et al.

Pre-trained foundation models (FMs) have shown exceptional performance in univariate time series forecasting tasks. However, several practical challenges persist, including managing intricate dependencies among features and quantifying uncertainty in predictions. This study aims to tackle these critical limitations by introducing adapters; feature-space transformations that facilitate the effective use of pre-trained univariate time series FMs for multivariate tasks. Adapters operate by projecting multivariate inputs into a suitable latent space and applying the FM independently to each dimension. Inspired by the literature on representation learning and partially stochastic Bayesian neural networks, we present a range of adapters and optimization/inference strategies. Experiments conducted on both synthetic and real-world datasets confirm the efficacy of adapters, demonstrating substantial enhancements in forecasting accuracy and uncertainty quantification compared to baseline methods. Our framework, AdaPTS, positions adapters as a modular, scalable, and effective solution for leveraging time series FMs in multivariate contexts, thereby promoting their wider adoption in real-world applications. We release the code at https://github.com/abenechehab/AdaPTS.

Abdelhakim Benechehab, Youssef Attia El Hili, Ambroise Odonnat et al.

The emerging zero-shot capabilities of Large Language Models (LLMs) have led to their applications in areas extending well beyond natural language processing tasks. In reinforcement learning, while LLMs have been extensively used in text-based environments, their integration with continuous state spaces remains understudied. In this paper, we investigate how pre-trained LLMs can be leveraged to predict in context the dynamics of continuous Markov decision processes. We identify handling multivariate data and incorporating the control signal as key challenges that limit the potential of LLMs' deployment in this setup and propose Disentangled In-Context Learning (DICL) to address them. We present proof-of-concept applications in two reinforcement learning settings: model-based policy evaluation and data-augmented off-policy reinforcement learning, supported by theoretical analysis of the proposed methods. Our experiments further demonstrate that our approach produces well-calibrated uncertainty estimates. We release the code at https://github.com/abenechehab/dicl.

Abdelhakim Benechehab, Giuseppe Paolo, Albert Thomas et al.

In model-based reinforcement learning (MBRL), most algorithms rely on simulating trajectories from one-step dynamics models learned on data. A critical challenge of this approach is the compounding of one-step prediction errors as length of the trajectory grows. In this paper we tackle this issue by using a multi-timestep objective to train one-step models. Our objective is a weighted sum of a loss function (e.g., negative log-likelihood) at various future horizons. We explore and test a range of weights profiles. We find that exponentially decaying weights lead to models that significantly improve the long-horizon R2 score. This improvement is particularly noticeable when the models were evaluated on noisy data. Finally, using a soft actor-critic (SAC) agent in pure batch reinforcement learning (RL) and iterated batch RL scenarios, we found that our multi-timestep models outperform or match standard one-step models. This was especially evident in a noisy variant of the considered environment, highlighting the potential of our approach in real-world applications.

Abdelhakim Benechehab, Gabriel Singer, Corentin Léger et al.

Generative models form the backbone of modern machine learning, underpinning state-of-the-art systems in text, vision, and multimodal applications. While Maximum Likelihood Estimation has traditionally served as the dominant training paradigm, recent work have highlighted its limitations, particularly in generalization and susceptibility to catastrophic forgetting compared to Reinforcement Learning techniques, such as Policy Gradient methods. However, these approaches depend on explicit reward signals, which are often unavailable in practice, leaving open the fundamental problem of how to align generative models when only high-quality datasets are accessible. In this work, we address this challenge via a Bilevel Optimization framework, where the reward function is treated as the optimization variable of an outer-level problem, while a policy gradient objective defines the inner-level. We then conduct a theoretical analysis of this optimization problem in a tractable setting and extract insights that, as we demonstrate, generalize to applications such as tabular classification and model-based reinforcement learning. We release the code at https://github.com/abenechehab/nll_to_po .

Youssef Attia El Hili, Albert Thomas, Malik Tiomoko et al.

Model and hyperparameter selection are critical but challenging in machine learning, typically requiring expert intuition or expensive automated search. We investigate whether large language models (LLMs) can act as in-context meta-learners for this task. By converting each dataset into interpretable metadata, we prompt an LLM to recommend both model families and hyperparameters. We study two prompting strategies: (1) a zero-shot mode relying solely on pretrained knowledge, and (2) a meta-informed mode augmented with examples of models and their performance on past tasks. Across synthetic and real-world benchmarks, we show that LLMs can exploit dataset metadata to recommend competitive models and hyperparameters without search, and that improvements from meta-informed prompting demonstrate their capacity for in-context meta-learning. These results highlight a promising new role for LLMs as lightweight, general-purpose assistants for model selection and hyperparameter optimization.

Antoine Grosnit, Alexandre Maraval, Refinath S N et al.

Human expertise emerges through iterative cycles of interaction, reflection, and internal model updating, which are central to cognitive theories such as Kolb's experiential learning and Vygotsky's zone of proximal development. In contrast, current AI systems, particularly LLM agents, rely on static pre-training or rigid workflows, lacking mechanisms for continual adaptation. Recent studies identified early cognitive traits in LLM agents (reflection, revision, and self-correction) suggesting foundational elements of human-like experiential learning. Thus the key question: Can we design LLM agents capable of structured, cognitively grounded learning similar to human processes? In response, we propose a computational framework of Kolb's learning cycle with Vygotsky's ZPD for autonomous agents. Our architecture separates extrinsic (environment interaction) and intrinsic (internal reflection/abstraction) functions, enabling cognitively grounded scaffolded learning, where the agent initially learns within structured environments, followed by open-ended generalisation. This approach empowers agents to master complex tasks ; domains that traditional fine-tuning or simple reflective methods could not tackle effectively. Its potential is powerfully demonstrated via direct comparison with humans in real-world Kaggle data science competitions. Learning fully automated data science code generation across 81 tasks, our system, Agent K, demonstrated the ability to perform the entire workflow autonomously, achieving an Elo-MMR score of 1694, beyond median score of the Kaggle Masters (the top 2% among 200,000 users) of our study. With 9 gold, 8 silver, and 12 bronze medals level performance - including 4 gold and 4 silver on prize-awarding competitions - Agent K is the 1st AI system to successfully integrate Kolb- and Vygotsky-inspired human cognitive learning, marking a major step toward generalist AI.

Abdelhakim Benechehab, Albert Thomas, Giuseppe Paolo et al.

In model-based reinforcement learning, most algorithms rely on simulating trajectories from one-step models of the dynamics learned on data. A critical challenge of this approach is the compounding of one-step prediction errors as the length of the trajectory grows. In this paper we tackle this issue by using a multi-step objective to train one-step models. Our objective is a weighted sum of the mean squared error (MSE) loss at various future horizons. We find that this new loss is particularly useful when the data is noisy (additive Gaussian noise in the observations), which is often the case in real-life environments. To support the multi-step loss, first we study its properties in two tractable cases: i) uni-dimensional linear system, and ii) two-parameter non-linear system. Second, we show in a variety of tasks (environments or datasets) that the models learned with this loss achieve a significant improvement in terms of the averaged R2-score on future prediction horizons. Finally, in the pure batch reinforcement learning setting, we demonstrate that one-step models serve as strong baselines when dynamics are deterministic, while multi-step models would be more advantageous in the presence of noise, highlighting the potential of our approach in real-world applications.

Giuseppe Paolo, Abdelhakim Benechehab, Hamza Cherkaoui et al.

Hierarchical organization is fundamental to biological systems and human societies, yet artificial intelligence systems often rely on monolithic architectures that limit adaptability and scalability. Current hierarchical reinforcement learning (HRL) approaches typically restrict hierarchies to two levels or require centralized training, which limits their practical applicability. We introduce TAME Agent Framework (TAG), a framework for constructing fully decentralized hierarchical multi-agent systems. TAG enables hierarchies of arbitrary depth through a novel LevelEnv concept, which abstracts each hierarchy level as the environment for the agents above it. This approach standardizes information flow between levels while preserving loose coupling, allowing for seamless integration of diverse agent types. We demonstrate the effectiveness of TAG by implementing hierarchical architectures that combine different RL agents across multiple levels, achieving improved performance over classical multi-agent RL baselines on standard benchmarks. Our results show that decentralized hierarchical organization enhances both learning speed and final performance, positioning TAG as a promising direction for scalable multi-agent systems.

Abdelhakim Benechehab, Albert Thomas, Balázs Kégl

We consider the problem of offline reinforcement learning where only a set of system transitions is made available for policy optimization. Following recent advances in the field, we consider a model-based reinforcement learning algorithm that infers the system dynamics from the available data and performs policy optimization on imaginary model rollouts. This approach is vulnerable to exploiting model errors which can lead to catastrophic failures on the real system. The standard solution is to rely on ensembles for uncertainty heuristics and to avoid exploiting the model where it is too uncertain. We challenge the popular belief that we must resort to ensembles by showing that better performance can be obtained with a single well-calibrated autoregressive model on the D4RL benchmark. We also analyze static metrics of model-learning and conclude on the important model properties for the final performance of the agent.

Alexandre Rio, Merwan Barlier, Igor Colin et al.

We address private deep offline reinforcement learning (RL), where the goal is to train a policy on standard control tasks that is differentially private (DP) with respect to individual trajectories in the dataset. To achieve this, we introduce PriMORL, a model-based RL algorithm with formal differential privacy guarantees. PriMORL first learns an ensemble of trajectory-level DP models of the environment from offline data. It then optimizes a policy on the penalized private model, without any further interaction with the system or access to the dataset. In addition to offering strong theoretical foundations, we demonstrate empirically that PriMORL enables the training of private RL agents on offline continuous control tasks with deep function approximations, whereas current methods are limited to simpler tabular and linear Markov Decision Processes (MDPs). We furthermore outline the trade-offs involved in achieving privacy in this setting.

Balázs Kégl, Gabriel Hurtado, Albert Thomas

We contribute to micro-data model-based reinforcement learning (MBRL) by rigorously comparing popular generative models using a fixed (random shooting) control agent. We find that on an environment that requires multimodal posterior predictives, mixture density nets outperform all other models by a large margin. When multimodality is not required, our surprising finding is that we do not need probabilistic posterior predictives: deterministic models are on par, in fact they consistently (although non-significantly) outperform their probabilistic counterparts. We also found that heteroscedasticity at training time, perhaps acting as a regularizer, improves predictions at longer horizons. At the methodological side, we design metrics and an experimental protocol which can be used to evaluate the various models, predicting their asymptotic performance when using them on the control problem. Using this framework, we improve the state-of-the-art sample complexity of MBRL on Acrobot by two to four folds, using an aggressive training schedule which is outside of the hyperparameter interval usually considered

Geovani Rizk, Igor Colin, Albert Thomas et al.

Experimental design is an approach for selecting samples among a given set so as to obtain the best estimator for a given criterion. In the context of linear regression, several optimal designs have been derived, each associated with a different criterion: mean square error, robustness, \emph{etc}. Computing such designs is generally an NP-hard problem and one can instead rely on a convex relaxation that considers probability distributions over the samples. Although greedy strategies and rounding procedures have received a lot of attention, straightforward sampling from the optimal distribution has hardly been investigated. In this paper, we propose theoretical guarantees for randomized strategies on E and G-optimal design. To this end, we develop a new concentration inequality for the eigenvalues of random matrices using a refined version of the intrinsic dimension that enables us to quantify the performance of such randomized strategies. Finally, we evidence the validity of our analysis through experiments, with particular attention on the G-optimal design applied to the best arm identification problem for linear bandits.

Geovani Rizk, Albert Thomas, Igor Colin et al.

We introduce a new graphical bilinear bandit problem where a learner (or a \emph{central entity}) allocates arms to the nodes of a graph and observes for each edge a noisy bilinear reward representing the interaction between the two end nodes. We study the best arm identification problem in which the learner wants to find the graph allocation maximizing the sum of the bilinear rewards. By efficiently exploiting the geometry of this bandit problem, we propose a \emph{decentralized} allocation strategy based on random sampling with theoretical guarantees. In particular, we characterize the influence of the graph structure (e.g. star, complete or circle) on the convergence rate and propose empirical experiments that confirm this dependency.

Igor Colin, Albert Thomas, Moez Draief

As cellular networks become denser, a scalable and dynamic tuning of wireless base station parameters can only be achieved through automated optimization. Although the contextual bandit framework arises as a natural candidate for such a task, its extension to a parallel setting is not straightforward: one needs to carefully adapt existing methods to fully leverage the multi-agent structure of this problem. We propose two approaches: one derived from a deterministic UCB-like method and the other relying on Thompson sampling. Thanks to its bayesian nature, the latter is intuited to better preserve the exploration-exploitation balance in the bandit batch. This is verified on toy experiments, where Thompson sampling shows robustness to the variability of the contexts. Finally, we apply both methods on a real base station network dataset and evidence that Thompson sampling outperforms both manual tuning and contextual UCB.

Stephan Clémençon, Albert Thomas

This paper aims at formulating the issue of ranking multivariate unlabeled observations depending on their degree of abnormality as an unsupervised statistical learning task. In the 1-d situation, this problem is usually tackled by means of tail estimation techniques: univariate observations are viewed as all the more `abnormal' as they are located far in the tail(s) of the underlying probability distribution. It would be desirable as well to dispose of a scalar valued `scoring' function allowing for comparing the degree of abnormality of multivariate observations. Here we formulate the issue of scoring anomalies as a M-estimation problem by means of a novel functional performance criterion, referred to as the Mass Volume curve (MV curve in short), whose optimal elements are strictly increasing transforms of the density almost everywhere on the support of the density. We first study the statistical estimation of the MV curve of a given scoring function and we provide a strategy to build confidence regions using a smoothed bootstrap approach. Optimization of this functional criterion over the set of piecewise constant scoring functions is next tackled. This boils down to estimating a sequence of empirical minimum volume sets whose levels are chosen adaptively from the data, so as to adjust to the variations of the optimal MV curve, while controling the bias of its approximation by a stepwise curve. Generalization bounds are then established for the difference in sup norm between the MV curve of the empirical scoring function thus obtained and the optimal MV curve.

Albert Thomas, Vincent Feuillard, Alexandre Gramfort

A general approach for anomaly detection or novelty detection consists in estimating high density regions or Minimum Volume (MV) sets. The One-Class Support Vector Machine (OCSVM) is a state-of-the-art algorithm for estimating such regions from high dimensional data. Yet it suffers from practical limitations. When applied to a limited number of samples it can lead to poor performance even when picking the best hyperparameters. Moreover the solution of OCSVM is very sensitive to the selection of hyperparameters which makes it hard to optimize in an unsupervised setting. We present a new approach to estimate MV sets using the OCSVM with a different choice of the parameter controlling the proportion of outliers. The solution function of the OCSVM is learnt on a training set and the desired probability mass is obtained by adjusting the offset on a test set to prevent overfitting. Models learnt on different train/test splits are then aggregated to reduce the variance induced by such random splits. Our approach makes it possible to tune the hyperparameters automatically and obtain nested set estimates. Experimental results show that our approach outperforms the standard OCSVM formulation while suffering less from the curse of dimensionality than kernel density estimates. Results on actual data sets are also presented.