Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Jae-Yun Jun has a total of 17 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Towards Language-Augmented Multi-Agent Deep Reinforcement Learning

    by Maxime Toquebiau, Jae-Yun Jun, Faïz Benamar, Nicolas Bredeche

    Most prior works on communication in multi-agent reinforcement learning have focused on emergent communication, which often results in inefficient and non-interpretable systems. Inspired by the role of language in natural intelligence, we investigate how grounding agents in a human-defined language can improve the learning and coordination of embodied agents. We propose a framework in which agents are trained not only to act but also to produce and interpret natural language descriptions of their observations. This language-augmented learning serves a dual role: enabling efficient and interpretable communication between agents, and guiding representation learning. We demonstrate that language-augmented agents outperform emergent communication baselines across various tasks. Our analysis reveals that language grounding leads to more informative internal representations, better generalization to new partners, and improved capability for human-agent interaction. These findings demonstrate the effectiveness of integrating structured language into multi-agent learning and open avenues for more interpretable and capable multi-agent systems.1 1 Appendices and code are available at github.com/MToquebiau/LAMARL


  • Open Access

    Approximate Closed-Form Solutions for Pricing Zero-Coupon Bonds in the Zero Lower Bound Framework

    by Jae-Yun Jun, Yves Rakotondratsimba

    After the 2007 financial crisis, many central banks adopted policies to lower their interest rates; the dynamics of these rates cannot be captured using classical models. Recently, Meucci and Loregian proposed an approach to estimate nonnegative interest rates using the inverse-call transformation. Despite the fact that their work is distinguished from others in the literature by their consideration of practical aspects, some technical difficulties still remain, such as the lack of analytic expression for the zero-coupon bond (ZCB) price. In this work, we propose novel approximate closed-form solutions for the ZCB price in the zero lower bound (ZLB) framework, when the underlying shadow rate is assumed to follow the classical one-factor Vasicek model. Then, a filtering procedure is performed using the Unscented Kalman Filter (UKF) to estimate the unobservable state variable (the shadow rate), and the model calibration is proceeded by estimating the model parameters using the Particle Swarm Optimization (PSO) algorithm. Further, empirical illustrations are given and discussed using (as input data) the interest rates of the AAA-rated bonds compiled by the European Central Bank ranging from 6 September 2004 to 21 June 2012 (a period that concerns the ZLB framework). Our approximate closed-form solution is able to show a good match between the actual and estimated yield-rate values for short and medium time-to-maturity values, whereas, for long time-to-maturity values, it is able to estimate the trend of the yield rates.

  • Open Access

    Joint Intrinsic Motivation for Coordinated Exploration in Multi-Agent Deep Reinforcement Learning

    by Maxime Toquebiau, Nicolas Bredeche, Faïz Benamar, Jae-Yun Jun

    Multi-agent deep reinforcement learning (MADRL) problems often encounter the challenge of sparse rewards. This challenge becomes even more pronounced when coordination among agents is necessary. As performance depends not only on one agent's behavior but rather on the joint behavior of multiple agents, finding an adequate solution becomes significantly harder. In this context, a group of agents can benefit from actively exploring different joint strategies in order to determine the most efficient one. In this paper, we propose an approach for rewarding strategies where agents collectively exhibit novel behaviors. We present JIM (Joint Intrinsic Motivation), a multi-agent intrinsic motivation method that follows the centralized learning with decentralized execution paradigm. JIM rewards joint trajectories based on a centralized measure of novelty designed to function in continuous environments. We demonstrate the strengths of this approach both in a synthetic environment designed to reveal shortcomings of state-of-the-art MADRL methods, and in simulated robotic tasks. Results show that joint exploration is crucial for solving tasks where the optimal strategy requires a high level of coordination.

  • Open Access

    A Multi-Agent Intrusion Detection System Optimized by a Deep Reinforcement Learning Approach with a Dataset Enlarged Using a Generative Model to Reduce the Bias Effect

    by Matthieu Mouyart, Guilherme Medeiros Machado, Jae-Yun Jun

    Intrusion detection systems can defectively perform when they are adjusted with datasets that are unbalanced in terms of attack data and non-attack data. Most datasets contain more non-attack data than attack data, and this circumstance can introduce biases in intrusion detection systems, making them vulnerable to cyberattacks. As an approach to remedy this issue, we considered the Conditional Tabular Generative Adversarial Network (CTGAN), with its hyperparameters optimized using the tree-structured Parzen estimator (TPE), to balance an insider threat tabular dataset called the CMU-CERT, which is formed by discrete-value and continuous-value columns. We showed through this method that the mean absolute errors between the probability mass functions (PMFs) of the actual data and the PMFs of the data generated using the CTGAN can be relatively small. Then, from the optimized CTGAN, we generated synthetic insider threat data and combined them with the actual ones to balance the original dataset. We used the resulting dataset for an intrusion detection system implemented with the Adversarial Environment Reinforcement Learning (AE-RL) algorithm in a multi-agent framework formed by an attacker and a defender. We showed that the performance of detecting intrusions using the framework of the CTGAN and the AE-RL is significantly improved with respect to the case where the dataset is not balanced, giving an F1-score of 0.7617.


  • Open Access

    Making green growth a reality: Reconciling sobriety with stakeholders’ satisfaction

    by Caroline Gans-Combe, Jae-Yun Jun, Waleed Mouhali, Yves Rakotondratsimba, Aïcha Baccar

    The notion of sobriety is considered a key variable in various energy transition scenarios. Often associated with a form of punitive ecology, it is, nevertheless, possible to make it a component that supports green growth, by linking it to the concept of "satisfaction". In this work, we have invented a way to achieve both “digital”, “economic”, and “ecological” sobriety, while ensuring the satisfaction of the end user. Directly correlated to the production of goods or services, the satisfaction function is built on the well-documented marginal utility function, which measures the need (or not) to consume further resources to satisfy the economic agents. Hence, it is justified and exists because it stands for the expectations of end users and makes sure the latter is met. This product itself is a function of the allocation of a set of resources, mapped using activity-based costing tools (ABC method). In this work, we focus on an AI proof-of-concept and demonstrate that it is possible to reach numerical sobriety by controlling the size of the training dataset while ensuring roughly the same model performance. In general, we show that it is possible to preserve the efficiency of AI processes while significantly minimizing the need for resources. In this sense, after establishing an analytical model, we suggest reducing the amount of data required to train the machine learning (ML) models, while guaranteeing zero change in terms of performance (say their accuracy). We show that it affects the energy consumed, and, thereby, the associated cost (i.e., economic and ecological) and the associated CO2eq emission. We thus confirm the existence of a "triangle of sobriety". It is defined as a virtual circle governed by a digital-economic-ecological sovereignty. We also propose that if AI production processes have a potential for sobriety, all identical activities have the same characteristics, thus opening the path to green growth.


  • APPROXIMATING OPTION PRICES UNDER LARGE CHANGES OF UNDERLYING ASSET PRICES

    by Jae-Yun Jun, Yves Rakotondratsimba

    When one invests in portfolios of derivatives (such as options), the delta-gamma approximation (DGA) is often used as a risk management strategy to reduce the risk associated with the underlying asset price. However, this approximation is locally accepted only for small changes of the underlying asset price. When these changes become large, the option prices estimated by the DGA may significantly differ from those of the market (or those that are estimated using, for instance, the Black–Scholes model), depending mainly on the time-to-maturity, implied volatility, and moneyness. Hence, in practice, before the change of the underlying asset price becomes large, rebalancing operations are demanded to minimize the losses occurred due to the error introduced by the DGA. The frequency of rebalancing may be high when the rate at which the underlying asset price significantly changes. Nonetheless, frequent rebalancing may be unattainable, as there are associated transaction costs. Hence, there is a trade-off between the losses resulting from the inaccurate performance of the DGA and the transaction costs incurring from frequent hedging operations. In the present work, we show two approaches that can outperform the DGA, in this way to increase the accuracy of estimating the option prices with the ultimate goal of reducing the losses due to the estimation error. The first method is similar to the DGA but we change the reference value that the DGA uses (that is, the initial price of the underlying asset) to the underlying asset price forecasted for the time horizon. We coin this method as the extended delta-gamma approximation (EDGA). The second method that we consider in this work is the locally weighted regression (LWR) that locally regresses the option prices from the changes of the underlying asset prices, with the same reference value that is employed in the EDGA method. Finally, we compare the performance of the two methods presented in this work to that of some existing methods.


  • Forecasting negative yield‐curve distributions

    by Jae‐Yun Jun, Victor Lebreton, Yves Rakotondratsimba

    Negative interest rates have been present in various marketplaces since mid‐2014, following the negative interest rate policy (NIRP) adopted by the European Central Bank to raise economic growth. The well‐known historical approach (HA) appears to be a good resource. By tweaking the HA, we derive a very tractable data‐driven tool that allows practitioners to generate yield‐curve distributions at future discrete time horizons. We thereby provide a robust and easy‐to‐understand forecasting model, suitable for the NIRP context, allowing an appreciation of its predictive power. Besides the methodological development herein, various numerical illustrations are presented to shed light on the benefits (and limitations) of this forecasting approach.

  • Characterization of running with compliant curved legs

    by Jae-Yun Jun, Jonathan E Clark

    Running with compliant curved legs involves the progression of the center of pressure, the changes of both the leg's stiffness and effective rest length, and the shift of the location of the maximum stress point along the leg. These phenomena are product of the geometric and material properties of these legs, and the rolling motion produced during stance. We examine these aspects with several reduced-order dynamical models to relate the leg's design parameters (such as normalized foot radius, leg's effective stiffness, location of the maximum stress point and leg shape) to running performance (such as robustness and efficiency). By using these models, we show that running with compliant curved legs can be more efficient, robust with fast recovery behavior from perturbations than running with compliant straight legs. Moreover, the running performance can be further improved by tuning these design parameters in the context of running with rolling. The results shown in this work may serve as potential guidance for future compliant curved leg designs that may further improve the running performance.


  • Open Access

    A trajectory tracking control design for a skid-steering mobile robot by adapting its desired instantaneous center of rotation

    by Jae-Yun Jun, Minh-Duc Hua, Faiz Benamar

    A skid-steering mobile robot steers by creating a moment that is larger than the frictional moment which results in a lateral slippage also known as skidding. This moment is in turn generated by a difference of the forces originated from the two sides of the robot. Tracking a given trajectory using this type of steering mechanism is not easy since it requires to relate skidding to steering. A necessary condition for the stability of skid-steering mobile robots is that the longitudinal component of the instantaneous center of rotation (ICR) resides within the robot dimension. In the present work, we propose a novel trajectory-tracking control design using a backstepping technique that guarantees the Lyapunov stability and that satisfies this necessary condition by relating the longitudinal component of the “desired ICR” to the curvature of a given trajectory and the reference linear speed. Finally, we compare the performance of the proposed controller to that of other existing controllers for skid-steering mobile robots and show the robustness of the proposed controller even in the presence of modeled sensory noise and control time delay in simulation.