Reinforcement Learning
Reinforcement Learning (RL) is a type of machine learning in which an autonomous agent learns to make sequential decisions by interacting with an environment. Through trial and error, the agent receives feedback in the form of rewards or penalties, allowing it to improve its strategy over time to maximize long-term outcomes. Unlike supervised learning, RL does not rely on labeled datasets—instead, it discovers optimal behaviors by exploring and evaluating the consequences of its actions.

At NEC Labs America, reinforcement learning is being applied to cutting-edge domains such as real-time camera optimization, drug discovery, and industrial efficiency—often in combination with imitation learning, adversarial learning, and other hybrid approaches to solve complex, real-world challenges.

Read our reinforcement learning publications below.

Posts

EditGRPO: Reinforcement Learning with Post-Rollout Edits for Clinically Accurate Chest X-Ray Report Generation

/in /by

Radiology report generation requires advanced medical image analysis, effective temporal reasoning, and accurate text generation. Although recent innovations, particularly multimodal large language models, have shown improved performance, their supervised fine-tuning (SFT) objective is not explicitly aligned with clinical efficacy. In this work, we introduce EditGRPO, a mixed-policy reinforcement learning algorithm designed specifically to optimize the generation through clinically motivated rewards. EditGRPO integrates on-policy exploration with off-policy guidance by injecting sentence-level detailed corrections during training rollouts. This mixed-policy approach addresses the exploration dilemma and sampling efficiency issues typically encountered in RL. Applied to a Qwen2.5-VL-3B, EditGRPO outperforms both SFT and vanilla GRPO baselines, achieving an average improvement of 3.4% in clinical metrics across four major datasets. Notably, EditGRPO also demonstrates superior out-of-domain generalization, with an average performance gain of5.9% on unseen datasets.

Read more

Visual Alignment of Medical Vision-Language Models for Grounded Radiology Report Generation

/in /by

Radiology Report Generation (RRG) is a critical step toward automating healthcare workflows, facilitating accurate patient assessments, and reducing the workload of medical professionals. Despite recent progress in Large Medical Vision-Language Models (Med-VLMs), generating radiology reports that are both visually grounded and clinically accurate remains a significant challenge. Existing approaches often rely on large labeled corpora for pre-training, costly task-specific preference data, or retrieval-based methods. However, these strategies do not adequately mitigate hallucinations arising from poor cross-modal alignment between visual and linguistic representations. To address these limitations, we propose VALOR:Visual Alignment of Medical Vision-Language Models for GrOunded Radiology Report Generation. Our method introduces a reinforcement learning-based post-alignment framework utilizing Group-Relative Proximal Optimization (GRPO). The training proceeds in two stages: (1) improving the Med-VLM with textual rewards to encourage clinically precise terminology, and (2) aligning the vision projection module of the textually grounded model with disease findings, thereby guiding attention toward image re gions most relevant to the diagnostic task. Extensive experiments on multiple benchmarks demonstrate that VALOR substantially improves factual accuracy and visual grounding, achieving significant performance gains over state-of-the-art report generation methods.

Read more

CAMTUNER: Adaptive Video Analytics Pipelines via Real-time Automated Camera Parameter Tuning

/in /by

In Video Analytics Pipelines (VAP), Analytics Units (AUs) such as object detection and face recognition operating on remote servers rely heavily on surveillance cameras to capture high-quality video streams to achieve high accuracy. Modern network cameras offer an array of parameters that directly influence video quality. While a few of such parameters, e.g., exposure, focus and white balance, are automatically adjusted by the camera internally, the others are not. We denote such camera parameters as non-automated (NAUTO) parameters. In this work, we first show that in a typical surveillance camera deployment, environmental condition changes can have significant adverse effect on the accuracy of insights from the AUs, but such adverse impact can potentially be mitigated by dynamically adjusting NAUTO camera parameters in response to changes in environmental conditions. Second, since most end-users lack the skill or understanding to appropriately configure these parameters and typically use a fixed parameter setting, we present CAMTUNER, to our knowledge, the first framework that dynamically adapts NAUTO camera parameters to optimize the accuracy of AUs in a VAP in response to adverse changes in environmental conditions. CAMTUNER is based on SARSA reinforcement learning and it incorporates two novel components: a light-weight analytics quality estimator and a virtual camera that drastically speed up offline RL training. Our controlled experiments and real-world VAP deployment show that compared to a VAP using the default camera setting, CAMTUNER enhances VAP accuracy by detecting 15.9% additional persons and 2.6%-4.2% additional cars (without any false positives) in a large enterprise parking lot. CAMTUNER opens up new avenues for elevating video analytics accuracy, transcending mere incremental enhancements achieved through refining deep-learning models.

Read more

PAIL: Performance based Adversarial Imitation Learning Engine for Carbon Neutral Optimization

/in /by

Achieving carbon neutrality within industrial operations has become increasingly imperative for sustainable development. It is both a significant challenge and a key opportunity for operational optimization in industry 4.0. In recent years, Deep Reinforcement Learning (DRL) based methods offer promising enhancements for sequential optimization processes and can be used for reducing car-bon emissions. However, existing DRL methods need a pre-defined reward function to assess the impact of each action on the final sustainable development goals (SDG). In many real applications, such a reward function cannot be given in advance. To address the problem, this study proposes a Performance based Adversarial Imitation Learning (PAIL) engine. It is a novel method to acquire optimal operational policies for carbon neutrality without any pre-defined action rewards. Specifically, PAIL employs a Transformer-based policy generator to encode historical information and predict fol-lowing actions within a multi-dimensional space. The entire action sequence will be iteratively updated by an environmental simulator. Then PAIL uses a discriminator to minimize the discrepancy be-tween generated sequences and real-world samples of high SDG. In parallel, a Q-learning framework based performance estimator is de-signed to estimate the impact of each action on SDG. Based on these estimations, PAIL refines generated policies with the rewards from both discriminator and performance estimator. PAIL is evaluated on multiple real-world application cases and datasets. The experiment results demonstrate the effectiveness of PAIL comparing to other state-of-the-art baselines. In addition, PAIL offers meaningful interpretability for the optimization in carbon neutrality.

Read more

Optimizing LLM API usage costs with novel query-aware reduction of relevant enterprise data

/in /by

Costs of LLM API usage rise rapidly when proprietary enterprise data is used as context for user queries to generate more accurate responses from LLMs. To reduce costs, we propose LeanContext, which generates query-aware, compact and AI model-friendly summaries of relevant enterprise data context. This is unlike traditional summarizers that produce query-unaware human-friendly summaries that are also not as compact. We first use retrieval augmented generation (RAG) to generate a query-aware enterprise data context, which includes key, query-relevant enterprise data. Then, we use reinforcement learning to further reduce the context while ensuring that a prompt consisting of the user query and the reduced context elicits an LLM response that is just as accurate as the LLM response to a prompt that uses the original enterprise data context. Our reduced context is not only query-dependent, but it is also variable-sized. Our experimental results demonstrate that LeanContext (a) reduces costs of LLM API usage by 37% to 68% (compared to RAG), while maintaining the accuracy of the LLM response, and (b) improves accuracy of responses by 26% to 38% when state-of-the-art summarizers reduce RAG context.

Read more

Advancing Sustainability in Global Supply Chains through Agent-based Simulation

/in /by

In today’s world, with its complex global supply chains, the difficulties and uncertainties we face offer both challenges and opportunities for making things better, especially in terms of efficiency and sustainability. These challenges grow due to unpredictable events, such as natural disasters, unexpected incidents, and unusual business practices, pushing us towards more advanced modeling methods that focus on reducing risks and enhancing sustainability. In this paper, we present a new agent-based simulation approach that goes beyond the usual limits of supply chain simulations by incorporating sustainability directly into supply chain operations using reinforcement learning (RL) algorithms. We introduce MOGI, a sustainable supply chain simulation system that takes carbon emissions into account in its main operations. Additionally, we examine how effective a multi-agent RL strategy is in dealing with the complex and uncertain nature of supply chains that span multiple levels. By comparing this strategy with traditional heuristic methods, our study looks at how well single versus multiple RL agents can manage risks and improve sustainability in both the beginning and end parts of the supply chain. The results of our experiments show that strategies based on RL are much better than traditional methods at managing risks, making profits, and achieving sustainability goals.

Read more

CLAP: Cost and Latency-Aware Placement of Microservices on the Computing Continuum

/in /by

For microservices-based real-time stream processing applications, computing at the edge delivers fast responses for low workloads, but as workload increases, the response time starts to slow down due to limited compute capacity. Abundant compute capacity in the cloud delivers fast responses even for higher workloads but incurs very high cost of operation. For applications which can tolerate latencies up to a certain limit, using either of them has one or the other drawback and for different applications and edge infrastructures, it is non-trivial to decide when to use only edge resources and when to leverage cloud resources. In this paper, we propose CLAP, which dynamically understands the relationship between workload and application latency, and automatically adjusts placement of microservices across edge and cloud computing continuum, with the goal of jointly reducing latency as well as cost of running microservices based streaming applications. CLAP leverages Reinforcement Learning (RL) technique to learn the optimal placement for a given workload and based on the learnings, adjusts placement of microservices as the application workload changes. We conduct experiments with real-world video analytics applications and show that CLAP adapts placement of microservices in response to varying workloads and achieves low latency for applications in a cost-efficient manner. Particularly, we show that for two real world video analytics applications i.e. human attributes and face recognition, CLAP is able to reduce average cost (across 4 days at different locations) by 47% and 58% for human attributes detection and face recognition application, respectively, while consistently maintaining latency below the tolerable limit.

Read more

Dynamic Causal Discovery in Imitation Learning

/in /by

Imitation learning, which learns agent policy by mimicking expert demonstration, has shown promising results in many applications such as medical treatment regimes and self-driving vehicles. However, it remains a difficult task to interpret control policies learned by the agent. Difficulties mainly come from two aspects: 1) agents in imitation learning are usually implemented as deep neural networks, which are black-box models and lack interpretability; 2) the latent causal mechanism behind agents’ decisions may vary along the trajectory, rather than staying static throughout time steps. To increase transparency and offer better interpretability of the neural agent, we propose to expose its captured knowledge in the form of a directed acyclic causal graph, with nodes being action and state variables and edges denoting the causal relations behind predictions. Furthermore, we design this causal discovery process to be state-dependent, enabling it to model the dynamics in latent causal graphs. Concretely, we conduct causal discovery from the perspective of Granger causality and propose a self-explainable imitation learning framework, CAIL. The proposed framework is composed of three parts: a dynamic causal discovery module, a causality encoding module, and a prediction module, and is trained in an end-to-end manner. After the model is learned, we can obtain causal relations among states and action variables behind its decisions, exposing policies learned by it. Experimental results on both synthetic and real-world datasets demonstrate the effectiveness of the proposed CAIL in learning the dynamic causal graphs for understanding the decision-making of imitation learning meanwhilemaintaining high prediction accuracy.

Read more

Calibrate Graph Neural Networks under Out-of-Distribution Nodes via Deep Q-learning

/in /by

Graph neural networks (GNNs) have achieved great success in dealing with graph-structured data that are prevalent in the real world. The core of graph neural networks is the message passing mechanism that aims to generate the embeddings of nodes by aggregating the neighboring node information. However, recent work suggests that GNNs also suffer the trustworthiness issues. Our empirical study shows that the calibration error of the in-distribution (ID) nodes would be exacerbated if a graph is mixed with out-of-distribution (OOD) nodes, and we assume that the noisy information from OOD nodes is the root for the worsened calibration error. Both previous study and our empirical study suggest that adjusting the weights of edges could be a promising way to reduce the adverse impact from the OOD nodes. However, how to precisely select the desired edges and modify the corresponding weights is not trivial, since the distribution of OOD nodes is unknown to us. To tackle this problem, we propose a Graph Edge Re-weighting via Deep Q-learning (GERDQ) framework to calibrate the graph neural networks. Our framework aims to explore the potential influence of the change of the edge weights on target ID nodes by sampling and traversing the edges in the graph, and we formulate this process as a Markov Decision Process (MDP). Many existing GNNs could be seamlessly incorporated into our framework. Experimental results show that when wrapped with our method, the existing GNN models can yield lower calibration error under OOD nodes as well as comparable accuracy compared to the original ones and other strong baselines. The source code is available at:https://github.com/DamoSWL/Calibration-GNN-OOD.

Read more

Few-Shot Video Classification via Representation Fusion and Promotion Learning

/in /by

Recent few-shot video classification (FSVC) works achieve promising performance by capturing similarity across support and query samples with different temporal alignment strategies or learning discriminative features via Transformer block within each episode. However, they ignore two important issues: a) It is difficult to capture rich intrinsic action semantics from a limited number of support instances within each task. b) Redundant or irrelevant frames in videos easily weaken the positive influence of discriminative frames. To address these two issues, this paper proposes a novel Representation Fusion and Promotion Learning (RFPL) mechanism with two sub-modules: meta-action learning (MAL) and reinforced image representation (RIR). Concretely, during training stage, we perform online learning for seeking a task-shared meta-action bank to enrich task-specific action representation by injecting global knowledge. Besides, we exploit reinforcement learning to obtain the importance of each frame and refine the representation. This operation maximizes the contribution of discriminative frames to further capture the similarity of support and query samples from the same category. Our RFPL framework is highly flexible that it can be integrated with many existing FSVC methods. Extensive experiments show that RFPL significantly enhances the performance of existing FSVC models when integrated with them.

Read more