Phillip Swazinna

Phillip Swazinna, Steffen Udluft, Thomas Runkler

Offline reinforcement learning algorithms still lack trust in practice due to the risk that the learned policy performs worse than the original policy that generated the dataset or behaves in an unexpected way that is unfamiliar to the user. At the same time, offline RL algorithms are not able to tune their most important hyperparameter - the proximity of the learned policy to the original policy. We propose an algorithm that allows the user to tune this hyperparameter at runtime, thereby addressing both of the above mentioned issues simultaneously. This allows users to start with the original behavior and grant successively greater deviation, as well as stopping at any time when the policy deteriorates or the behavior is too far from the familiar one.

Marc Weber, Phillip Swazinna, Daniel Hein et al.

Offline reinforcement learning provides a viable approach to obtain advanced control strategies for dynamical systems, in particular when direct interaction with the environment is not available. In this paper, we introduce a conceptual extension for model-based policy search methods, called variable objective policy (VOP). With this approach, policies are trained to generalize efficiently over a variety of objectives, which parameterize the reward function. We demonstrate that by altering the objectives passed as input to the policy, users gain the freedom to adjust its behavior or re-balance optimization targets at runtime, without need for collecting additional observation batches or re-training.

Phillip Swazinna, Steffen Udluft, Thomas Runkler

Recently, offline RL algorithms have been proposed that remain adaptive at runtime. For example, the LION algorithm \cite{lion} provides the user with an interface to set the trade-off between behavior cloning and optimality w.r.t. the estimated return at runtime. Experts can then use this interface to adapt the policy behavior according to their preferences and find a good trade-off between conservatism and performance optimization. Since expert time is precious, we extend the methodology with an autopilot that automatically finds the correct parameterization of the trade-off, yielding a new algorithm which we term AutoLION.

Jean-Philippe Corbeil, Asma Ben Abacha, Jerome Tremblay et al.

Clinical documentation increasingly uses automatic speech recognition and summarization, yet converting conversations into actionable medical orders for Electronic Health Records remains unexplored. A solution to this problem can significantly reduce the documentation burden of clinicians and directly impact downstream patient care. We introduce the MEDIQA-OE 2025 shared task, the first challenge on extracting medical orders from doctor-patient conversations. Six teams participated in the shared task and experimented with a broad range of approaches, and both closed- and open-weight large language models (LLMs). In this paper, we describe the MEDIQA-OE task, dataset, final leaderboard ranking, and participants' solutions.

Jean-Philippe Corbeil, Asma Ben Abacha, George Michalopoulos et al. · microsoft-research

Large language models (LLMs) such as GPT-4o and o1 have demonstrated strong performance on clinical natural language processing (NLP) tasks across multiple medical benchmarks. Nonetheless, two high-impact NLP tasks - structured tabular reporting from nurse dictations and medical order extraction from doctor-patient consultations - remain underexplored due to data scarcity and sensitivity, despite active industry efforts. Practical solutions to these real-world clinical tasks can significantly reduce the documentation burden on healthcare providers, allowing greater focus on patient care. In this paper, we investigate these two challenging tasks using private and open-source clinical datasets, evaluating the performance of both open- and closed-weight LLMs, and analyzing their respective strengths and limitations. Furthermore, we propose an agentic pipeline for generating realistic, non-sensitive nurse dictations, enabling structured extraction of clinical observations. To support further research in both areas, we release SYNUR and SIMORD, the first open-source datasets for nurse observation extraction and medical order extraction.

Amna Najib, Stefan Depeweg, Phillip Swazinna

Batch reinforcement learning enables policy learning without direct interaction with the environment during training, relying exclusively on previously collected sets of interactions. This approach is, therefore, well-suited for high-risk and cost-intensive applications, such as industrial control. Learned policies are commonly restricted to act in a similar fashion as observed in the batch. In a real-world scenario, learned policies are deployed in the industrial system, inevitably leading to the collection of new data that can subsequently be added to the existing recording. The process of learning and deployment can thus take place multiple times throughout the lifespan of a system. In this work, we propose to exploit this iterative nature of applying offline reinforcement learning to guide learned policies towards efficient and informative data collection during deployment, leading to continuous improvement of learned policies while remaining within the support of collected data. We present an algorithmic methodology for iterative batch reinforcement learning based on ensemble-based model-based policy search, augmented with safety and, importantly, a diversity criterion.

Batıkan Bora Ormancı, Phillip Swazinna, Steffen Udluft et al.

In this paper, we investigate offline reinforcement learning (RL) with the goal of training a single robust policy that generalizes effectively across environments with unseen dynamics. We propose a novel approach, Trajectory Encoding Augmentation (TEA), which extends the state space by integrating latent representations of environmental dynamics obtained from sequence encoders, such as AutoEncoders. Our findings show that incorporating these encodings with TEA improves the transferability of a single policy to novel environments with new dynamics, surpassing methods that rely solely on unmodified states. These results indicate that TEA captures critical, environment-specific characteristics, enabling RL agents to generalize effectively across dynamic conditions.

Phillip Swazinna, Steffen Udluft, Daniel Hein et al.

Offline reinforcement learning (RL) Algorithms are often designed with environments such as MuJoCo in mind, in which the planning horizon is extremely long and no noise exists. We compare model-free, model-based, as well as hybrid offline RL approaches on various industrial benchmark (IB) datasets to test the algorithms in settings closer to real world problems, including complex noise and partially observable states. We find that on the IB, hybrid approaches face severe difficulties and that simpler algorithms, such as rollout based algorithms or model-free algorithms with simpler regularizers perform best on the datasets.

Phillip Swazinna, Steffen Udluft, Thomas Runkler

Recently developed offline reinforcement learning algorithms have made it possible to learn policies directly from pre-collected datasets, giving rise to a new dilemma for practitioners: Since the performance the algorithms are able to deliver depends greatly on the dataset that is presented to them, practitioners need to pick the right dataset among the available ones. This problem has so far not been discussed in the corresponding literature. We discuss ideas how to select promising datasets and propose three very simple indicators: Estimated relative return improvement (ERI) and estimated action stochasticity (EAS), as well as a combination of the two (COI), and empirically show that despite their simplicity they can be very effectively used for dataset selection.

Phillip Swazinna, Steffen Udluft, Daniel Hein et al.

In offline reinforcement learning, a policy needs to be learned from a single pre-collected dataset. Typically, policies are thus regularized during training to behave similarly to the data generating policy, by adding a penalty based on a divergence between action distributions of generating and trained policy. We propose a new algorithm, which constrains the policy directly in its weight space instead, and demonstrate its effectiveness in experiments.

Phillip Swazinna, Steffen Udluft, Thomas Runkler

State-of-the-art reinforcement learning algorithms mostly rely on being allowed to directly interact with their environment to collect millions of observations. This makes it hard to transfer their success to industrial control problems, where simulations are often very costly or do not exist, and exploring in the real environment can potentially lead to catastrophic events. Recently developed, model-free, offline RL algorithms, can learn from a single dataset (containing limited exploration) by mitigating extrapolation error in value functions. However, the robustness of the training process is still comparatively low, a problem known from methods using value functions. To improve robustness and stability of the learning process, we use dynamics models to assess policy performance instead of value functions, resulting in MOOSE (MOdel-based Offline policy Search with Ensembles), an algorithm which ensures low model bias by keeping the policy within the support of the data. We compare MOOSE with state-of-the-art model-free, offline RL algorithms { BRAC,} BEAR and BCQ on the Industrial Benchmark and MuJoCo continuous control tasks in terms of robust performance, and find that MOOSE outperforms its model-free counterparts in almost all considered cases, often even by far.