Author name cluster

Joschka Boedecker

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

27 papers

2 author rows

TMLR Journal 2026 Journal Article

Goal Achievement Guided Exploitation: A Principled Performance-Based Scheduling Framework for Reinforcement Learning

Shengchao Yan
Baohe Zhang
Chenguang Huang
Joschka Boedecker
Wolfram Burgard

In dense-reward tasks, Reinforcement learning (RL) algorithms often employ soft entropy regularization to promote exploration. By integrating an entropy term into the objective function, they regularize exploration via tuning the coefficient. However, the entropy coefficient only indirectly influences the action distribution through gradient updates, making it difficult to precisely control exploration, and requires careful scheduling to balance exploration and exploitation throughout training. As a solution, we propose Goal Achievement Guided Exploitation~(GAGE), a performance-based scheduling framework that adaptively regulates exploration by linking policy stochasticity directly to the agent's performance relative to a target value. Unlike soft entropy regularizers, GAGE enforces hard, performance-dependent constraints on action distribution's standard deviation for continuous actions and logit range for discrete actions. Consequently, GAGE ensures a guaranteed lower bound on action probabilities that naturally decays as the agent approaches optimal performance. Across a suite of challenging robotic control tasks, GAGE improves learning efficiency and stability across various strong baselines, achieving competitive or superior final performance. By providing a more interpretable and robust alternative to entropy-based exploration heuristics, GAGE offers a scalable path toward solving complex dense reward tasks with pronounced local optima.

PDF Details

TMLR Journal 2025 Journal Article

Constrained Reinforcement Learning with Smoothed Log Barrier Function

Baohe Zhang
Yuan Zhang
Hao Zhu
Shengchao Yan
Thomas Brox
Joschka Boedecker

Deploying reinforcement learning (RL) in real-world systems often requires satisfying strict safety constraints during both training and deployment, which simple reward shaping typically fails to enforce. Existing constrained RL algorithms frequently face several major challenges, including instabilities during training and overly conservative policies. To overcome these limitations, we propose CSAC-LB (Constrained Soft Actor-Critic with Log Barrier), a model-free, sample-efficient, off-policy algorithm that requires no pre-training. CSAC-LB integrates a linear smoothed log barrier function into the actor’s objective, providing a numerically stable, non-vanishing gradient that enables the agent to quickly recover from unsafe states while avoiding the instability of traditional interior-point methods. To further enhance safety and mitigate the underestimation of constraint violations, we employ a pessimistic double-critic architecture for the cost function, taking the maximum of two cost Q-networks to conservatively guide the policy. Through extensive experiments on challenging constrained control tasks, we demonstrate that CSAC-LB significantly outperforms baselines by consistently achieving high returns while strictly adhering to safety constraints. Our results establish CSAC-LB as a robust and stable solution for applying RL to safety-critical domains.

PDF Details

TMLR Journal 2025 Journal Article

Multi-BK-Net: Multi-Branch Multi-Kernel Convolutional Neural Networks for Clinical EEG Analysis

Ann-Kathrin Kiessner
Tonio Ball
Joschka Boedecker

Classifying an electroencephalography (EEG) recording as pathological or non-pathological is an important first step in diagnosing and managing neurological diseases and disorders. As manual EEG classification is costly, time-consuming and requires highly trained experts, deep learning methods for automated classification of general EEG pathology offer a promising option to assist clinicians in screening EEGs. Convolutional neural networks (CNNs) are well-suited for classifying pathological EEG signals due to their ability to perform end-to-end learning. In practice, however, current CNN solutions suffer from limited classification performance due to I) a single-scale network design that cannot fully capture the high intra- and inter-subject variability of the EEG signal, the diversity of the data, and the heterogeneity of pathological EEG patterns and II) the small size and limited diversity of the dataset commonly used to train and evaluate the networks. These challenges result in a low sensitivity score and a performance drop on more diverse patient populations, further hindering their reliability for real-world applications. Here, we propose a novel multi-branch, multi-scale CNN called Multi-BK-Net (Multi-Branch Multi-Kernel Network), comprising five parallel branches that incorporate temporal convolution, spatial convolution, and pooling layers, with temporal kernel sizes defined by five clinically relevant frequency bands in its first block. Evaluation is based on two public datasets with predefined test sets: the Temple University Hospital (TUH) Abnormal EEG Corpus and the TUH Abnormal Expansion Balanced EEG Corpus. Our Multi-BK-Net outperforms five baseline architectures and state-of-the-art end-to-end approaches in terms of accuracy and sensitivity on these datasets, setting a new benchmark. Furthermore, ablation experiments highlight the importance of the multi-branch, multi-scale input block of the Multi-BK-Net. Overall, our findings indicate the efficacy of multi-branch, multi-scale CNNs in accurately and reliably classifying EEG pathology, demonstrating advantages in handling data heterogeneity compared to other deep learning approaches. Thus, this study contributes to the ongoing development of deep end-to-end methods for general EEG pathology classification.

PDF Details

ICLR Conference 2025 Conference Paper

Salvage: Shapley-distribution Approximation Learning Via Attribution Guided Exploration for Explainable Image Classification

Mehdi Naouar
Hanne Raum
Jens Rahnfeld
Yannick Vogt
Joschka Boedecker
Gabriel Kalweit
Maria Kalweit

The integration of deep learning into critical vision application areas has given rise to a necessity for techniques that can explain the rationale behind predictions. In this paper, we address this need by introducing Salvage, a novel removal-based explainability method for image classification. Our approach involves training an explainer model that learns the prediction distribution of the classifier on masked images. We first introduce the concept of Shapley-distributions, which offers a more accurate approximation of classification probability distributions than existing methods. Furthermore, we address the issue of unbalanced important and unimportant features. In such settings, naive uniform sampling of feature subsets often results in a highly unbalanced ratio of samples with high and low prediction likelihoods, which can hinder effective learning. To mitigate this, we propose an informed sampling strategy that leverages approximated feature importance scores, thereby reducing imbalance and facilitating the estimation of underrepresented features. After incorporating these two principles into our method, we conducted an extensive analysis on the ImageNette, MURA, WBC, and Pet datasets. The results show that Salvage outperforms various baseline explainability methods, including attention-, gradient-, and removal-based approaches, both qualitatively and quantitatively. Furthermore, we demonstrate that our explainer model can serve as a fully explainable classifier without a major decrease in classification performance, paving the way for fully explainable image classification.