Category Archives: Reinforcement Learning In Ai

Explaining the results of RL in continuous spaces

September 3, 2026 09:35 ,

R. Zuo, S. Khan, Z. Wang, G. E. Katz and Q. Qiu, Beyond VisionMask: Explaining Reinforcement Learning Via Contrastive-State Masking, IEEE Intelligent Systems, vol. 41, no. 4, pp. 56-63, July-Aug. 2026,10.1109/MIS.2026.3683960.

Existing explainable reinforcement learning (RL) approaches often fail to capture the contrastive nature of human reasoning—answering “why this action instead of that one?”. To address this challenge, we present Continuous VisionMask (cVM), a contrastive learning framework that explains RL agents in continuous state and action spaces. cVM extends our prior work, VisionMask, which was restricted to vision-based discrete settings, by introducing a quantized representation of continuous state and a set of contrastive learning objectives. This generalization enables cVM to provide faithful, robust, and sparse attributions highlighting which aspects of the sensed input are most important for the agent’s decision making, and to do so across a broader range of real-world RL scenarios. We evaluate cVM in multiple continuous-control environments and compare it against existing explainability baselines. Our results demonstrate that cVM produces more faithful and stable explanations, thereby enhancing transparency and interpretability in continuous RL systems.

Posted in: Reinforcement learning in AI , Tagged:

Unlearning repeatitive cycles in RL

September 3, 2026 09:04 ,

Dinesh Kumar Banjare, Mithilesh Atulkar, Mitul Kumar Ahirwal, Utilizing machine unlearning for faulty behavior patterns in human decision-making model: A sharded reinforcement learning approach, Engineering Applications of Artificial Intelligence, Volume 181, Part 6, 2026,10.1016/j.engappai.2026.115652.

In this paper, enhance the computational models of human decision-making by proposing a novel sharded reinforcement learning framework based on the principle of machine unlearning. The work proposes a data sharding and selective machine unlearning approach to unlearn faulty behavior patterns. It overcomes the drawbacks of conventional reinforcement learning methods, in which the players tend to play the same deck over and over again, showing a lack of adaptability. These problems are frequently seen in tasks such as the Iowa gambling task. To handle this issue, the concept of machine unlearning is used. The proposed method detects repeated patterns of faulty behavior, such as repeated deck selections. Such patterns are corrected locally at the level of shards, while retaining the last stable expectancy state from the previous learning stage. The mean squared deviation was used to evaluate the model’s performance. The results indicate that the unlearned model learns faster and has better performance in the early phases of learning. The framework supports behavioral adaptability by selectively excluding identified faulty shards while preserving previous valid learning. These findings show the potential of shard-based learning and selective unlearning in sequential decision-making.

Posted in: Reinforcement learning in AI , Tagged:

Hierarchical adversarial RL

September 3, 2026 08:19 ,

Xiaofeng Wang, Shaopu Wang, Jipeng Zhang, Yue Ma, Hierarchical multi-policy adversarial reinforcement learning, Engineering Applications of Artificial Intelligence, Volume 181, Part 1, 2026, 10.1016/j.engappai.2026.115297.

In complex and dynamic autonomous driving environments, deep reinforcement learning policies often struggle with generalization due to distribution shifts between training and deployment. While conventional approaches like domain randomization and adversarial training offer partial solutions, they face notable limitations. Specifically, domain randomization relies on manually defined distributions that poorly capture real-world perturbations, while adversarial training suffers from high computational costs, unstable optimization, and behavioral homogenization. To address these issues, we propose a hierarchical multi-policy adversarial reinforcement learning framework that enhances policy generalization under complex perturbations. It incorporates three key components: a multi-policy generation architecture with a shared backbone and independent heads for diverse adversary modeling; a dual-perspective policy expansion regularization mechanism; and an adaptive curriculum-based adversarial training framework. The engineering value translates to a reliable, high-performance solution for autonomous driving, as validated on high-fidelity driving benchmarks. Experimentally, our approach demonstrates superior convergence, stability, and out-of-distribution generalization, achieving performance improvements ranging from 7.36% to 50.06% over established baseline methods while maintaining computational efficiency. These results, obtained in a high-fidelity simulator, demonstrate promise for real-world autonomous driving applications and lay the groundwork for future real-world deployment.

Posted in: Reinforcement learning in AI , Tagged: ,

Selecting the best RL result from offline RL in order to minimize risks

April 10, 2026 16:58 ,

Giorgio Angelotti, Nicolas Drougard, Caroline P․ C. Chanel, An offline risk-aware policy selection method for Bayesian Markov decision processes, Artificial Intelligence, Volume 354, 2026, 10.1016/j.artint.2026.104519.

In Offline Model Learning for Planning and in Offline Reinforcement Learning, the limited data set hinders the estimate of the Value function of the relative Markov Decision Process (MDP). Consequently, the performance of the obtained policy in the real world is bounded and possibly risky, especially when the deployment of a wrong policy can lead to catastrophic consequences. For this reason, several pathways are being followed with the scope of reducing the model error (or the distributional shift between the learned model and the true one) and, more broadly, obtaining risk-aware solutions with respect to model uncertainty. But when it comes to the final application which baseline should a practitioner choose? In an offline context where computational time is not an issue and robustness is the priority we propose Exploitation vs Caution (EvC), a paradigm that (1) elegantly incorporates model uncertainty abiding by the Bayesian formalism, and (2) selects the policy that maximizes a risk-aware objective over the Bayesian posterior between a fixed set of candidate policies provided, for instance, by the current baselines. We validate EvC with state-of-the-art approaches in different discrete, yet simple, environments offering a fair variety of MDP classes. In the tested scenarios EvC manages to select robust policies and hence stands out as a useful tool for practitioners that aim to apply offline planning and reinforcement learning solvers in the real world.

Posted in: Reinforcement learning in AI , Tagged:

Using evolutionary computation to find better rewards in the case of partial-observable RL

February 19, 2026 10:26 ,

Zhengwei Zhu, Zhixuan Chen, Chenyang Zhu, Wen Si, Fang Wang, Optimizing potential-based reward automata in partially observable reinforcement learning using genetic local search, Engineering Applications of Artificial Intelligence, Volume 169, 2026, 10.1016/j.engappai.2026.114054.

Partially observable reinforcement learning extends the reinforcement learning framework to environments in which agents have limited visibility of the state space, making it particularly relevant for applications in robotics and autonomous vehicle navigation. However, a primary challenge in partially observable reinforcement learning is defining effective reward functions that can guide the learning process despite partial observability. To address this challenge, this paper introduces a novel approach for constructing potential-based reward automata by employing genetic local search methods. Specifically, our method constructs these automata from compressed representations of exploration trajectories, which succinctly capture critical decision points and essential state transitions while eliminating redundant steps. By optimizing trajectory samples and shortening agent trajectories to their crucial transitions, our technique significantly reduces computational overhead. Formally, we define the learning objective as an optimization problem aimed at maximizing the log-likelihood of future observations while simultaneously minimizing the structural complexity of the learned reward automata. Furthermore, by incorporating value-based strategies to estimate potential values within the reward automata, our approach improves learning efficiency and facilitates the identification of optimal reward structures. We empirically evaluate our proposed method on seven partially observable grid-world benchmarks. Experimental results demonstrate that our method achieves superior performance relative to state-of-the-art reward automata-based techniques, exhibiting both accelerated learning speeds and higher accumulated rewards. Additionally, our genetic local search algorithm consistently outperforms comparative heuristic methods in terms of learning curves and reward accumulation.

Posted in: Reinforcement learning in AI , Tagged: , ,

RL with both discrete and continuous actions

November 6, 2025 15:27 ,

Chengcheng Yan, Shujie Chen, Jiawei Xu, Xuejie Wang, Zheng Peng, Hybrid Reinforcement Learning in parameterized action space via fluctuates constraint, Engineering Applications of Artificial Intelligence, Volume 162, Part C, 2025 10.1016/j.engappai.2025.112499.

Parameterized actions in Reinforcement Learning (RL) are composed of discrete-continuous hybrid action parameters, which are widely employed in game scenarios. However, previous works have often concentrated on the network structure of RL algorithms to solve hybrid actions, neglecting the impact of fluctuations in action parameters for agent move trajectory. Due to the coupling between discrete and continuous actions, instability in discrete actions influences the selection of corresponding continuous parameters, resulting in the agent deviating from the optimal move path. In this paper, we propose a parameterized RL approach based on parameter fluctuation restriction (PFR) to address this problem, called CP-DQN. Our method effectively mitigated value fluctuation in action parameters by constraining the action parameter between adjacent time steps. Additionally, we have incorporated a supervision module to optimize the entire training process. To quantify the superiority of our approach in minimizing trajectory deviations for agents, we propose an indicator to measure the influence of parameter fluctuations on performance in hybrid action space. Our method is evaluated in three environments with hybrid action spaces, and the experiments demonstrate the superiority of our method compared to existing approaches.

Posted in: Reinforcement learning in AI

A variant of RL aimed at reducing bias of conventional Q-learning

November 6, 2025 15:23 ,

Fanghui Huang, Wenqi Han, Xiang Li, Xinyang Deng, Wen Jiang, Reducing the estimation bias and variance in reinforcement learning via Maxmean and Aitken value iteration, Engineering Applications of Artificial Intelligence, Volume 162, Part C, 2025, 10.1016/j.engappai.2025.112502.

The value-based reinforcement leaning methods suffer from overestimation bias, because of the existence of max operator, resulting in suboptimal policies. Meanwhile, variance in value estimation will cause the instability of networks. Many algorithms have been presented to solve the mentioned, but these lack the theoretical analysis about the degree of estimation bias, and the trade-off between the estimation bias and variance. Motivated by the above, in this paper, we propose a novel method based on Maxmean and Aitken value iteration, named MMAVI. The Maxmean operation allows the average of multiple state–action values (Q values) to be used as the estimated target value to mitigate the bias and variance. The Aitken value iteration is used to update Q values and improve the convergence rate. Based on the proposed method, combined with Q-learning and deep Q-network, we design two novel algorithms to adapt to different environments. To understand the effect of MMAVI, we analyze it both theoretically and empirically. In theory, we derive the closed-form expressions of reducing bias and variance, and prove that the convergence rate of our proposed method is faster than the traditional methods with Bellman equation. In addition, the convergence of our algorithms is proved in a tabular setting. Finally, we demonstrate that our proposed algorithms outperform the state-of-the-art algorithms in several environments.

Posted in: Reinforcement learning in AI , Tagged:

A quantitative demonstration based on MDPs of the increasing need of a world model (learnt or given) as the complexity of the task and the performance of the agent increase

October 23, 2025 11:44 ,

Jonathan Richens, David Abel, Alexis Bellot, Tom Everitt, General agents contain world models, arXiv cs:AI, Sep. 2025, arXiv:2506.01622.

Are world models a necessary ingredient for flexible, goal-directed behaviour, or is model-free learning sufficient? We provide a formal answer to this question, showing that any agent capable of generalizing to multi-step goal-directed tasks must have learned a predictive model of its environment. We show that this model can be extracted from the agent’s policy, and that increasing the agents performance or the complexity of the goals it can achieve requires learning increasingly accurate world models. This has a number of consequences: from developing safe and general agents, to bounding agent capabilities in complex environments, and providing new algorithms for eliciting world models from agents.

Posted in: Reinforcement learning in AI , Tagged: , , ,

Inclusion of LLMs in multiple task learning for generating rewards

October 16, 2025 08:30 ,

Z. Lin, Y. Chen and Z. Liu, AutoSkill: Hierarchical Open-Ended Skill Acquisition for Long-Horizon Manipulation Tasks via Language-Modulated Rewards, IEEE Transactions on Cognitive and Developmental Systems, vol. 17, no. 5, pp. 1141-1152, Oct. 2025, 10.1109/TCDS.2025.3551298.

A desirable property of generalist robots is the ability to both bootstrap diverse skills and solve new long-horizon tasks in open-ended environments without human intervention. Recent advancements have shown that large language models (LLMs) encapsulate vast-scale semantic knowledge about the world to enable long-horizon robot planning. However, they are typically restricted to reasoning high-level instructions and lack world grounding, which makes it difficult for them to coordinately bootstrap and acquire new skills in unstructured environments. To this end, we propose AutoSkill, a hierarchical system that empowers the physical robot to automatically learn to cope with new long-horizon tasks by growing an open-ended skill library without hand-crafted rewards. AutoSkill consists of two key components: 1) an in-context skill chain generation and new skill bootstrapping guided by LLMs that inform the robot of discrete and interpretable skill instructions for skill retrieval and augmentation within the skill library; and 2) a zero-shot language-modulated reward scheme in conjunction with a meta prompter facilitates online new skill acquisition via expert-free supervision aligned with proposed skill directives. Extensive experiments conducted in both simulated and realistic environments demonstrate AutoSkill’s superiority over other LLM-based planners as well as hierarchical methods in expediting online learning for novel manipulation tasks.

Posted in: Reinforcement learning in AI , Tagged: , ,

Improvements in offline RL (from previously acquired datasets)

October 9, 2025 08:51 ,

Lan Wu, Quan Liu, Renyang You, State slow feature softmax Q-value regularization for offline reinforcement learning, Engineering Applications of Artificial Intelligence, Volume 160, Part A, 2025, 10.1016/j.engappai.2025.111828.

Offline reinforcement learning is constrained by its reliance on pre-collected datasets, without the opportunity for further interaction with the environment. This restriction often results in distribution shifts, which can exacerbate Q-value overestimation and degrade policy performance. To address these issues, we propose a method called state slow feature softmax Q-value regularization (SQR), which enhances the stability and accuracy of Q-value estimation in offline settings. SQR employs slow feature representation learning to extract dynamic information from state trajectories, promoting the stability and robustness of the state representations. Additionally, a softmax operator is incorporated into the Q-value update process to smooth Q-value estimation, reducing overestimation and improving policy optimization. Finally, we apply our approach to locomotion and navigation tasks and establish a comprehensive experimental analysis framework. Empirical results demonstrate that SQR outperforms state-of-the-art offline RL baselines, achieving performance improvements ranging from 2.5% to 44.6% on locomotion tasks and 2.0% to 71.1% on navigation tasks. Moreover, it achieves the highest score on 7 out of 15 locomotion datasets and 4 out of 6 navigation datasets. Detailed experimental results confirm the stabilizing effect of slow feature learning and the effectiveness of the softmax regularization in mitigating Q-value overestimation, demonstrating the superiority of SQR in addressing key challenges in offline reinforcement learning.

Posted in: Reinforcement learning in AI , Tagged: