Emmanuel Rachelson

David Bertoin, Adil Zouitine, Mehdi Zouitine et al.

Deep reinforcement learning policies, despite their outstanding efficiency in simulated visual control tasks, have shown disappointing ability to generalize across disturbances in the input training images. Changes in image statistics or distracting background elements are pitfalls that prevent generalization and real-world applicability of such control policies. We elaborate on the intuition that a good visual policy should be able to identify which pixels are important for its decision, and preserve this identification of important sources of information across images. This implies that training of a policy with small generalization gap should focus on such important pixels and ignore the others. This leads to the introduction of saliency-guided Q-networks (SGQN), a generic method for visual reinforcement learning, that is compatible with any value function learning method. SGQN vastly improves the generalization capability of Soft Actor-Critic agents and outperforms existing stateof-the-art methods on the Deepmind Control Generalization benchmark, setting a new reference in terms of training efficiency, generalization gap, and policy interpretability.

Valentin Guillet, Dennis G. Wilson, Carlos Aguilar-Melchor et al.

Learning a good state representation is a critical skill when dealing with multiple tasks in Reinforcement Learning as it allows for transfer and better generalization between tasks. However, defining what constitute a useful representation is far from simple and there is so far no standard method to find such an encoding. In this paper, we argue that distillation -- a process that aims at imitating a set of given policies with a single neural network -- can be used to learn a state representation displaying favorable characteristics. In this regard, we define three criteria that measure desirable features of a state encoding: the ability to select important variables in the input space, the ability to efficiently separate states according to their corresponding optimal action, and the robustness of the state encoding on new tasks. We first evaluate these criteria and verify the contribution of distillation on state representation on a toy environment based on the standard inverted pendulum problem, before extending our analysis on more complex visual tasks from the Atari and Procgen benchmarks.

David Bertoin, Emmanuel Rachelson

Over the past few years, the acceleration of computing resources and research in deep learning has led to significant practical successes in a range of tasks, including in particular in computer vision. Building on these advances, reinforcement learning has also seen a leap forward with the emergence of agents capable of making decisions directly from visual observations. Despite these successes, the over-parametrization of neural architectures leads to memorization of the data used during training and thus to a lack of generalization. Reinforcement learning agents based on visual inputs also suffer from this phenomenon by erroneously correlating rewards with unrelated visual features such as background elements. To alleviate this problem, we introduce a new regularization technique consisting of channel-consistent local permutations (CLOP) of the feature maps. The proposed permutations induce robustness to spatial correlations and help prevent overfitting behaviors in RL. We demonstrate, on the OpenAI Procgen Benchmark, that RL agents trained with the CLOP method exhibit robustness to visual changes and better generalization properties than agents trained using other state-of-the-art regularization techniques. We also demonstrate the effectiveness of CLOP as a general regularization technique in supervised learning.

Paul Strang, Zacharie Alès, Côme Bissuel et al.

Mixed-Integer Linear Programming (MILP) lies at the core of many real-world combinatorial optimization (CO) problems, traditionally solved by branch-and-bound (B&B). A key driver influencing B&B solvers efficiency is the variable selection heuristic that guides branching decisions. Looking to move beyond static, hand-crafted heuristics, recent work has explored adapting traditional reinforcement learning (RL) algorithms to the B&B setting, aiming to learn branching strategies tailored to specific MILP distributions. In parallel, RL agents have achieved remarkable success in board games, a very specific type of combinatorial problems, by leveraging environment simulators to plan via Monte Carlo Tree Search (MCTS). Building on these developments, we introduce Plan-and-Branch-and-Bound (PlanB&B), a model-based reinforcement learning (MBRL) agent that leverages a learned internal model of the B&B dynamics to discover improved branching strategies. Computational experiments empirically validate our approach, with our MBRL branching agent outperforming previous state-of-the-art RL methods across four standard MILP benchmarks.

David Bertoin, Eduardo Hugo Sanchez, Mehdi Zouitine et al.

Despite transformers being considered as the new standard in computer vision, convolutional neural networks (CNNs) still outperform them in low-data regimes. Nonetheless, CNNs often make decisions based on narrow, specific regions of input images, especially when training data is limited. This behavior can severely compromise the model's generalization capabilities, making it disproportionately dependent on certain features that might not represent the broader context of images. While the conditions leading to this phenomenon remain elusive, the primary intent of this article is to shed light on this observed behavior of neural networks. Our research endeavors to prioritize comprehensive insight and to outline an initial response to this phenomenon. In line with this, we introduce Saliency Guided Dropout (SGDrop), a pioneering regularization approach tailored to address this specific issue. SGDrop utilizes attribution methods on the feature map to identify and then reduce the influence of the most salient features during training. This process encourages the network to diversify its attention and not focus solely on specific standout areas. Our experiments across several visual classification benchmarks validate SGDrop's role in enhancing generalization. Significantly, models incorporating SGDrop display more expansive attributions and neural activity, offering a more comprehensive view of input images in contrast to their traditionally trained counterparts.

Valentin Guillet, Dennis G. Wilson, Carlos Aguilar-Melchor et al.

Although transfer learning is considered to be a milestone in deep reinforcement learning, the mechanisms behind it are still poorly understood. In particular, predicting if knowledge can be transferred between two given tasks is still an unresolved problem. In this work, we explore the use of network distillation as a feature extraction method to better understand the context in which transfer can occur. Notably, we show that distillation does not prevent knowledge transfer, including when transferring from multiple tasks to a new one, and we compare these results with transfer without prior distillation. We focus our work on the Atari benchmark due to the variability between different games, but also to their similarities in terms of visual features.

Paul-Antoine Le Tolguenec, Emmanuel Rachelson, Yann Besse et al.

When searching for policies, reward-sparse environments often lack sufficient information about which behaviors to improve upon or avoid. In such environments, the policy search process is bound to blindly search for reward-yielding transitions and no early reward can bias this search in one direction or another. A way to overcome this is to use intrinsic motivation in order to explore new transitions until a reward is found. In this work, we use a recently proposed definition of intrinsic motivation, Curiosity, in an evolutionary policy search method. We propose Curiosity-ES, an evolutionary strategy adapted to use Curiosity as a fitness metric. We compare Curiosity with Novelty, a commonly used diversity metric, and find that Curiosity can generate higher diversity over full episodes without the need for an explicit diversity criterion and lead to multiple policies which find reward.

Thomas Delliaux, Nguyen-Khanh Vu, Vincent François-Lavet et al.

Learning meaningful abstract models of Markov Decision Processes (MDPs) is crucial for improving generalization from limited data. In this work, we show how geometric priors can be imposed on the low-dimensional representation manifold of a learned transition model. We incorporate known symmetric structures via appropriate choices of the latent space and the associated group actions, which encode prior knowledge about invariances in the environment. In addition, our framework allows the embedding of additional unstructured information alongside these symmetries. We show experimentally that this leads to better predictions of the latent transition model than fully unstructured approaches, as well as better learning on downstream RL tasks, in environments with rotational and translational features, including in first-person views of 3D environments. Additionally, our experiments show that this leads to simpler and more disentangled representations. The full code is available on GitHub to ensure reproducibility.

Adil Zouitine, David Bertoin, Pierre Clavier et al.

Robust reinforcement learning is the problem of learning control policies that provide optimal worst-case performance against a span of adversarial environments. It is a crucial ingredient for deploying algorithms in real-world scenarios with prevalent environmental uncertainties and has been a long-standing object of attention in the community, without a standardized set of benchmarks. This contribution endeavors to fill this gap. We introduce the Robust Reinforcement Learning Suite (RRLS), a benchmark suite based on Mujoco environments. RRLS provides six continuous control tasks with two types of uncertainty sets for training and evaluation. Our benchmark aims to standardize robust reinforcement learning tasks, facilitating reproducible and comparable experiments, in particular those from recent state-of-the-art contributions, for which we demonstrate the use of RRLS. It is also designed to be easily expandable to new environments. The source code is available at \href{https://github.com/SuReLI/RRLS}{https://github.com/SuReLI/RRLS}.

Paul Strang, Zacharie Alès, Côme Bissuel et al.

Mixed-Integer Linear Programming (MILP) is a powerful framework used to address a wide range of NP-hard combinatorial optimization problems, often solved by Branch and Bound (B&B). A key factor influencing the performance of B&B solvers is the variable selection heuristic governing branching decisions. Recent contributions have sought to adapt reinforcement learning (RL) algorithms to the B&B setting to learn optimal branching policies, through Markov Decision Processes (MDP) inspired formulations, and ad hoc convergence theorems and algorithms. In this work, we introduce BBMDP, a principled vanilla MDP formulation for variable selection in B&B, allowing to leverage a broad range of RL algorithms for the purpose of learning optimal B\&B heuristics. Computational experiments validate our model empirically, as our branching agent outperforms prior state-of-the-art RL agents on four standard MILP benchmarks.

Paul Strang, Zacharie Alès, Côme Bissuel et al.

On the occasion of the 20th Mixed Integer Program Workshop's computational competition, this work introduces a new approach for learning to solve MIPs online. Influence branching, a new graph-oriented variable selection strategy, is applied throughout the first iterations of the branch and bound algorithm. This branching heuristic is optimized online with Thompson sampling, which ranks the best graph representations of MIP's structure according to computational speed up over SCIP. We achieve results comparable to state of the art online learning methods. Moreover, our results indicate that our method generalizes well to more general online frameworks, where variations in constraint matrix, constraint vector and objective coefficients can all occur and where more samples are available.

Paul-Antoine Le Tolguenec, Yann Besse, Florent Teichteil-Koenigsbuch et al.

The ability to explore efficiently and effectively is a central challenge of reinforcement learning. In this work, we consider exploration through the lens of information theory. Specifically, we cast exploration as a problem of maximizing the Shannon entropy of the state occupation measure. This is done by maximizing a sequence of divergences between distributions representing an agent's past behavior and its current behavior. Intuitively, this encourages the agent to explore new behaviors that are distinct from past behaviors. Hence, we call our method RAMP, for ``$\textbf{R}$unning $\textbf{A}$way fro$\textbf{m}$ the $\textbf{P}$ast.'' A fundamental question of this method is the quantification of the distribution change over time. We consider both the Kullback-Leibler divergence and the Wasserstein distance to quantify divergence between successive state occupation measures, and explain why the former might lead to undesirable exploratory behaviors in some tasks. We demonstrate that by encouraging the agent to explore by actively distancing itself from past experiences, it can effectively explore mazes and a wide range of behaviors on robotic manipulation and locomotion tasks.

Paul-Antoine Le Tolguenec, Yann Besse, Florent Teichteil-Konigsbuch et al.

The ability to perform different skills can encourage agents to explore. In this work, we aim to construct a set of diverse skills which uniformly cover the state space. We propose a formalization of this search for diverse skills, building on a previous definition based on the mutual information between states and skills. We consider the distribution of states reached by a policy conditioned on each skill and leverage the successor state measure to maximize the difference between these skill distributions. We call this approach LEADS: Learning Diverse Skills through Successor States. We demonstrate our approach on a set of maze navigation and robotic control tasks which show that our method is capable of constructing a diverse set of skills which exhaustively cover the state space without relying on reward or exploration bonuses. Our findings demonstrate that this new formalization promotes more robust and efficient exploration by combining mutual information maximization and exploration bonuses.

Adil Zouitine, David Bertoin, Pierre Clavier et al.

Robust reinforcement learning is essential for deploying reinforcement learning algorithms in real-world scenarios where environmental uncertainty predominates. Traditional robust reinforcement learning often depends on rectangularity assumptions, where adverse probability measures of outcome states are assumed to be independent across different states and actions. This assumption, rarely fulfilled in practice, leads to overly conservative policies. To address this problem, we introduce a new time-constrained robust MDP (TC-RMDP) formulation that considers multifactorial, correlated, and time-dependent disturbances, thus more accurately reflecting real-world dynamics. This formulation goes beyond the conventional rectangularity paradigm, offering new perspectives and expanding the analytical framework for robust RL. We propose three distinct algorithms, each using varying levels of environmental information, and evaluate them extensively on continuous control benchmarks. Our results demonstrate that these algorithms yield an efficient tradeoff between performance and robustness, outperforming traditional deep robust RL methods in time-constrained environments while preserving robustness in classical benchmarks. This study revisits the prevailing assumptions in robust RL and opens new avenues for developing more practical and realistic RL applications.

Pierre Clavier, Emmanuel Rachelson, Erwan Le Pennec et al.

Many classic Reinforcement Learning (RL) algorithms rely on a Bellman operator, which involves an expectation over the next states, leading to the concept of bootstrapping. To introduce a form of pessimism, we propose to replace this expectation with an expectile. In practice, this can be very simply done by replacing the $L_2$ loss with a more general expectile loss for the critic. Introducing pessimism in RL is desirable for various reasons, such as tackling the overestimation problem (for which classic solutions are double Q-learning or the twin-critic approach of TD3) or robust RL (where transitions are adversarial). We study empirically these two cases. For the overestimation problem, we show that the proposed approach, ExpectRL, provides better results than a classic twin-critic. On robust RL benchmarks, involving changes of the environment, we show that our approach is more robust than classic RL algorithms. We also introduce a variation of ExpectRL combined with domain randomization which is competitive with state-of-the-art robust RL agents. Eventually, we also extend \ExpectRL with a mechanism for choosing automatically the expectile value, that is the degree of pessimism

Kaitlin Maile, Emmanuel Rachelson, Hervé Luga et al.

Neurogenesis in ANNs is an understudied and difficult problem, even compared to other forms of structural learning like pruning. By decomposing it into triggers and initializations, we introduce a framework for studying the various facets of neurogenesis: when, where, and how to add neurons during the learning process. We present the Neural Orthogonality (NORTH*) suite of neurogenesis strategies, combining layer-wise triggers and initializations based on the orthogonality of activations or weights to dynamically grow performant networks that converge to an efficient size. We evaluate our contributions against other recent neurogenesis works across a variety of supervised learning tasks.

David Bertoin, Emmanuel Rachelson

Deep neural networks have demonstrated their ability to automatically extract meaningful features from data. However, in supervised learning, information specific to the dataset used for training, but irrelevant to the task at hand, may remain encoded in the extracted representations. This remaining information introduces a domain-specific bias, weakening the generalization performance. In this work, we propose splitting the information into a task-related representation and its complementary context representation. We propose an original method, combining adversarial feature predictors and cyclic reconstruction, to disentangle these two representations in the single-domain supervised case. We then adapt this method to the unsupervised domain adaptation problem, consisting of training a model capable of performing on both a source and a target domain. In particular, our method promotes disentanglement in the target domain, despite the absence of training labels. This enables the isolation of task-specific information from both domains and a projection into a common representation. The task-specific representation allows efficient transfer of knowledge acquired from the source domain to the target domain. In the single-domain case, we demonstrate the quality of our representations on information retrieval tasks and the generalization benefits induced by sharpened task-specific representations. We then validate the proposed method on several classical domain adaptation benchmarks and illustrate the benefits of disentanglement for domain adaptation.

Thibault Lahire, Matthieu Geist, Emmanuel Rachelson

Several algorithms have been proposed to sample non-uniformly the replay buffer of deep Reinforcement Learning (RL) agents to speed-up learning, but very few theoretical foundations of these sampling schemes have been provided. Among others, Prioritized Experience Replay appears as a hyperparameter sensitive heuristic, even though it can provide good performance. In this work, we cast the replay buffer sampling problem as an importance sampling one for estimating the gradient. This allows deriving the theoretically optimal sampling distribution, yielding the best theoretical convergence speed. Elaborating on the knowledge of the ideal sampling scheme, we exhibit new theoretical foundations of Prioritized Experience Replay. The optimal sampling distribution being intractable, we make several approximations providing good results in practice and introduce, among others, LaBER (Large Batch Experience Replay), an easy-to-code and efficient method for sampling the replay buffer. LaBER, which can be combined with Deep Q-Networks, distributional RL agents or actor-critic methods, yields improved performance over a diverse range of Atari games and PyBullet environments, compared to the base agent it is implemented on and to other prioritization schemes.

Erwan Lecarpentier, David Abel, Kavosh Asadi et al.

We consider the problem of knowledge transfer when an agent is facing a series of Reinforcement Learning (RL) tasks. We introduce a novel metric between Markov Decision Processes (MDPs) and establish that close MDPs have close optimal value functions. Formally, the optimal value functions are Lipschitz continuous with respect to the tasks space. These theoretical results lead us to a value-transfer method for Lifelong RL, which we use to build a PAC-MDP algorithm with improved convergence rate. Further, we show the method to experience no negative transfer with high probability. We illustrate the benefits of the method in Lifelong RL experiments.

Andrea Lodi, Luca Mossina, Emmanuel Rachelson

We present an approach to couple the resolution of Combinatorial Optimization problems with methods from Machine Learning, applied to the single source, capacitated, facility location problem. Our study is framed in the context where a reference facility location optimization problem is given. Assuming there exist data for many variations of the reference problem (historical or simulated) along with their optimal solution, we study how one can exploit these to make predictions about an unseen new instance. We demonstrate how a classifier can be built from these data to determine whether the solution to the reference problem still applies to a new instance. In case the reference solution is partially applicable, we build a regressor indicating the magnitude of the expected change, and conversely how much of it can be kept for the new instance. This insight, derived from a priori information, is expressed via an additional constraint in the original mathematical programming formulation. We present an empirical evaluation and discuss the benefits, drawbacks and perspectives of such an approach. Although presented through the application to the facility location problem, the approach developed here is general and explores a new perspective on the exploitation of past experience in combinatorial optimization.

Luca Mossina, Emmanuel Rachelson, Daniel Delahaye

We study how Reinforcement Learning can be employed to optimally control parameters in evolutionary algorithms. We control the mutation probability of a (1+1) evolutionary algorithm on the OneMax function. This problem is modeled as a Markov Decision Process and solved with Value Iteration via the known transition probabilities. It is then solved via Q-Learning, a Reinforcement Learning algorithm, where the exact transition probabilities are not needed. This approach also allows previous expert or empirical knowledge to be included into learning. It opens new perspectives, both formally and computationally, for the problem of parameter control in optimization.

Erwan Lecarpentier, Emmanuel Rachelson

This work tackles the problem of robust zero-shot planning in non-stationary stochastic environments. We study Markov Decision Processes (MDPs) evolving over time and consider Model-Based Reinforcement Learning algorithms in this setting. We make two hypotheses: 1) the environment evolves continuously with a bounded evolution rate; 2) a current model is known at each decision epoch but not its evolution. Our contribution can be presented in four points. 1) we define a specific class of MDPs that we call Non-Stationary MDPs (NSMDPs). We introduce the notion of regular evolution by making an hypothesis of Lipschitz-Continuity on the transition and reward functions w.r.t. time; 2) we consider a planning agent using the current model of the environment but unaware of its future evolution. This leads us to consider a worst-case method where the environment is seen as an adversarial agent; 3) following this approach, we propose the Risk-Averse Tree-Search (RATS) algorithm, a zero-shot Model-Based method similar to Minimax search; 4) we illustrate the benefits brought by RATS empirically and compare its performance with reference Model-Based algorithms.

Erwan Lecarpentier, Guillaume Infantes, Charles Lesire et al.

In the context of tree-search stochastic planning algorithms where a generative model is available, we consider on-line planning algorithms building trees in order to recommend an action. We investigate the question of avoiding re-planning in subsequent decision steps by directly using sub-trees as action recommender. Firstly, we propose a method for open loop control via a new algorithm taking the decision of re-planning or not at each time step based on an analysis of the statistics of the sub-tree. Secondly, we show that the probability of selecting a suboptimal action at any depth of the tree can be upper bounded and converges towards zero. Moreover, this upper bound decays in a logarithmic way between subsequent depths. This leads to a distinction between node-wise optimality and state-wise optimality. Finally, we empirically demonstrate that our method achieves a compromise between loss of performance and computational gain.

Luca Mossina, Emmanuel Rachelson

This article focuses on the question of learning how to automatically select a subset of items among a bigger set. We introduce a methodology for the inference of ensembles of discrete values, based on the Naive Bayes assumption. Our motivation stems from practical use cases where one wishes to predict an unordered set of (possibly interdependent) values from a set of observed features. This problem can be considered in the context of Multi-label Classification (MLC) where such values are seen as labels associated to continuous or discrete features. We introduce the \nbx algorithm, an extension of Naive Bayes classification into the multi-label domain, discuss its properties and evaluate our approach on real-world problems.

Erwan Lecarpentier, Sebastian Rapp, Marc Melo et al.

Autonomous unpowered flight is a challenge for control and guidance systems: all the energy the aircraft might use during flight has to be harvested directly from the atmosphere. We investigate the design of an algorithm that optimizes the closed-loop control of a glider's bank and sideslip angles, while flying in the lower convective layer of the atmosphere in order to increase its mission endurance. Using a Reinforcement Learning approach, we demonstrate the possibility for real-time adaptation of the glider's behaviour to the time-varying and noisy conditions associated with thermal soaring flight. Our approach is online, data-based and model-free, hence avoids the pitfalls of aerological and aircraft modelling and allow us to deal with uncertainties and non-stationarity. Additionally, we put a particular emphasis on keeping low computational requirements in order to make on-board execution feasible. This article presents the stochastic, time-dependent aerological model used for simulation, together with a standard aircraft model. Then we introduce an adaptation of a Q-learning algorithm and demonstrate its ability to control the aircraft and improve its endurance by exploiting updrafts in non-stationary scenarios.