Partially Observable RL

Many safety-critical control problems are modeled as risk-sensitive partially observable Markov decision processes, where the controller must make decisions from incomplete observations while balancing task performance against safety risk. Although belief-space planning provides a principled solution, maintaining and planning over beliefs can be computationally costly and sensitive to model specification in practical domains. We propose a lightweight risk-gated reinforcement learning approximation for risk-sensitive control under partial observability. The method constructs a compact finite-history proxy state and learns an action-conditioned predictor of near-term safety violation. This predicted candidate-action risk is used in two complementary ways: as a risk penalty during value learning, and as a decision-time gate that interpolates between optimistic and conservative ensemble value estimates. As a result, low-risk actions are evaluated closer to reward-seeking estimates, while high-risk actions are evaluated more conservatively. We evaluate the approach in two safety-critical partially observable domains: automated glucose regulation and safety-constrained navigation. Across adult and adolescent glucose-control cohorts, the method improves overall glycemic tradeoffs and substantially reduces runtime relative to a belief-space planning baseline. On Safety-Gym navigation benchmarks, it achieves a more favorable reward-cost balance than unconstrained RL and several standard safe-RL baselines. These results suggest that action-conditioned near-term risk can provide an effective local signal for approximate risk-sensitive POMDP control when full belief-space planning is impractical.

In reinforcement learning (RL), agents acting in partially observable Markov decision processes (POMDPs) must rely on memory, typically encoded in a recurrent neural network (RNN), to integrate information from past observations. Long-horizon POMDPs, in which the relevant observation and the optimal action are separated by many time steps (called the horizon), are particularly challenging: training suffers from poor generalization, severe sample inefficiency, and prohibitive exploration costs. Ideally, an agent trained on short horizons would retain optimal behavior at arbitrarily longer ones, but no formal framework currently characterizes when this is achievable. To fill this gap, we formalized temporal horizon generalization, the property that a policy remains optimal for all horizons, derived a necessary and sufficient condition for it, and experimentally evaluated the ability of nonlinear and parallelizable RNN variants to achieve it. This paper presents the resulting theoretical framework, the empirical evaluation, and the dynamical interpretation linking RNN behavior to temporal horizon generalization. Our analyses reveal that multistability is necessary for temporal horizon generalization and, in simple tasks, sufficient; more complex tasks further require transient dynamics. In contrast, modern parallelizable architectures, namely state space models and gated linear RNNs, are monostable by construction and consequently fail to generalize across temporal horizons. We conclude that multistability and transient dynamics are two essential and complementary dynamical regimes for horizon generalization, and that no current parallelizable RNN exhibits both. Designing parallelizable architectures that combine these regimes thus emerges as a key direction for scalable long-horizon RL.

Task completion in digital and physical environments increasingly involves complex temporal interaction, where actions and observations unfold over different time scales rather than align with fixed observation--action steps. To model such interactions, we propose \emph{Engagement Process} (EP), an interaction formalism that inherits the decision-theoretic structure of POMDPs while making time explicit in the action--observation interface. EP represents actions and observations as decoupled event streams along time, rather than updates paired at fixed decision steps. This interface captures single-agent timing issues such as deliberation latency, delayed feedback, and persistent actions, while supporting richer agent-side organization, multi-rate coordination, and compositional interaction among subsystems. Across toy, LLM-agent, and learning experiments, EP exposes temporal behaviors hidden by step-based interfaces and enables policies to adapt under explicit time costs.

Large language models (LLMs) are increasingly deployed on long-horizon tasks in partially observable environments, where they must act while inferring and tracking a complex environment state over many steps. This leads to two challenges: partial observability requires maintaining uncertainty over unobserved world attributes, and long interaction history causes context to grow without bound, diluting task-relevant information. A principled solution to both challenges is a belief state: a posterior distribution over environment states given past observations and actions, which compactly encodes history for decision making regardless of episode length. In LLM agents, however, the open-ended nature of text makes it unclear how to represent such a distribution. Therefore, we introduce Agent-BRACE: Agent Belief state Representation via Abstraction and Confidence Estimation, a method that decouples an LLM agent into a belief state model and a policy model, jointly optimized via reinforcement learning. The belief state model produces a structured approximation of the belief distribution: a set of atomic natural language claims about the environment, each annotated with an ordinal verbalized certainty label ranging from certain to unknown. The policy model conditions on this compact, structured approximate belief rather than the full history, learning to select actions under explicit uncertainty. Across long-horizon, partially observable embodied language environments, Agent-BRACE achieves an average absolute improvement of +14.5% (Qwen2.5-3B-Instruct) and +5.3% (Qwen3-4B-Instruct), outperforming strong RL baselines while maintaining a near-constant context window independent of episode length. Further analysis shows that the learned belief becomes increasingly calibrated over the course of an episode as evidence accumulates.

Large language models (LLMs) have shown remarkable in-context learning (ICL) capabilities, yet their potential for sequential decision-making remains underexplored. In this paper, we study the ICL capabilities of LLMs in sequential decision-making settings, including Markov Decision Processes (MDPs), Partially Observable MDPs (POMDPs), and Ambiguous POMDPs (APOMDPs). We fine-tune pretrained LLMs to perform few-shot decision-making directly from offline, oracle-labeled trajectories. Our framework enables flexible imitation of policies through supervised fine-tuning (SFT). Theoretically, we focus on linear MDPs and interpret a fine-tuned attention layer as implicitly estimating optimal Q-functions from in-context data. Building on this interpretation, we derive an end-to-end suboptimality bound for the induced policy that separates the in-context estimation error from the training-length bias. Empirically, across synthetic MDP, POMDP, and APOMDP settings, we find that fine-tuned LLMs achieve substantially smaller optimality gaps than in-context-only and random baselines, with especially large gains in longer-horizon, partially observed, and model-ambiguous environments. Together, these results show that supervised fine-tuning provides an effective route to endowing pretrained LLMs with sequential decision-making capabilities from offline data, which is an important advantage in domains such as healthcare where offline data are abundant.

Cooperative multi-agent reinforcement learning agents that act on partial local observations face a fundamental information bottleneck: the knowledge needed to select jointly optimal actions is scattered across the team, yet each agent must commit to a decision without access to its teammates' observations, intentions, or chosen actions. Existing methods either ignore this bottleneck, compress it into a scalar mixing signal, or route around it with learned communication channels. Framing action coordination as a problem of structured information integration among agents, we propose \textit{structured agent coordination via holistic information integration}, or SACHI, in which graph transformer convolutions over an inter-agent coordination graph enrich each agent's representation with receiver-sensitive, content-dependent signals from teammates prior to action selection. We evaluate SACHI across five cooperative tasks spanning spatial, communicative, and adversarial coordination challenges against twelve baselines. SACHI consistently matches or outperforms the best baseline on every task, and rigorous aggregate statistical analyses, including normalized metrics with bootstrap confidence intervals, Friedman ranking, and performance profiling, confirm that this advantage is statistically significant, robust across environments, and not attributable to increased model capacity. Parameter-matched ablations further trace the source of the gains to a single architectural property: the degree of content-dependence in the message-passing operator.

Outcome metrics can certify the wrong behavior. We study this failure in a two-hotel revenue-management simulator where Hotel A trains an agent against a fixed rule-based revenue-management competitor, Hotel B. A standard learning agent can obtain near-reference revenue per available room (RevPAR) while failing to learn market-like yield management: it sells too aggressively, undercuts, or collapses to modal price buckets. We diagnose this as a Goodhart-style failure under partial observability. Hotel A cannot observe the competitor's remaining inventory, booking curve, or pricing rule, so the same Hotel A-visible state maps to multiple plausible Hotel B prices. Deterministic value-based RL and deterministic copying collapse this unresolved uncertainty into shortcut behavior. We introduce a trace-level diagnostic protocol using RevPAR, occupancy, ADR, full price-bucket distributions, L1/JS distances, and seed-level confidence intervals. The verified repair is Trace-Prior RL: learn a distributional market prior from lagged market traces, then train a stochastic pricing policy with a RevPAR reward and a KL penalty to the learned prior. The final policy matches Hotel B's RevPAR, occupancy, ADR, and price distribution within seed-level uncertainty, while still optimizing Hotel A's own reward. We argue that the contribution is not a new optimizer and not a hotel-pricing leaderboard, but a reproducible failure-and-repair recipe for agentic systems where scalar rewards are easy to game and the intended behavior is only visible in traces. A key finding is that higher exact action accuracy can worsen aggregate trace alignment when the target is distributional.

Causal reinforcement learning (RL) lacks benchmarks for complex systems that combine sequential decision making, hidden information, large masked action spaces, and explicit causal structure. We introduce MTG-Causal-RL, a Gymnasium benchmark built on Magic: The Gathering with a 3,077-dimensional partial observation, a 478-action masked discrete action space, five competitive Standard archetypes, three reward schemes, and a hand-specified Structural Causal Model (SCM) over strategic variables. Every episode exposes causal variables, SCM-predicted intervention effects, and per-factor credit traces, making causal credit assignment, leave-one-out cross-archetype transfer, and policy auditability first-class metrics. We adapt a panel of reference baselines: random, heuristic, masked PPO, a causal-world-model PPO variant, and an architecture-matched scalar control. We propose Causal Graph-Factored Advantage PPO (CGFA-PPO) as a reference causal agent that uses SCM parents of win probability as factor-aligned critic targets with an intervention-calibration loss. All comparisons use paired seeds, paired-bootstrap confidence intervals, and Holm-Bonferroni correction within pre-registered families. Masked PPO and CGFA-PPO reach competitive in-distribution win rates and exceed the random baseline; per-factor calibration trajectories and leave-one-out transfer gaps expose diagnostic structure that scalar win rate alone cannot. We release the benchmark, reference-baseline results, and full evaluation protocol openly. By coupling a strategically rich, partially observed domain with an explicit causal interface and statistical protocol, MTG-Causal-RL gives causal-RL, world-model, and LLM-agent research a shared testbed for questions current benchmarks cannot pose together: causal credit assignment under masked action spaces, structural transfer across archetypes, and SCM-grounded policy auditability.

A key capability of intelligent agents is to act effectively under incomplete state observations. Recurrent policies address this by compressing observation histories into a hidden state. In this work, we show that the hidden state of a recurrent policy admits a control-theoretic interpretation: it plays the role of the co-state in Pontryagin's minimum principle, and the readout that maps it to actions implements control-Hamiltonian minimization. The hidden state thus tracks the gradient of the value function, encoding the optimality structure of the underlying control problem. We formalize this correspondence through a class of policies we refer to as co-state policies (CPs) and show that several modern recurrent cells implicitly realize this structure. The correspondence also allows for a co-state loss for actor-critic training, in which the critic's gradient serves as a target for the actor's hidden state. Empirically, we find that hidden states trained with the co-state loss encode co-state information beyond what is linearly decodable from the environment state alone, and that the loss improves policies both within and outside this class on challenging locomotion tasks, including the H1 and Berkeley humanoids. By connecting the minimum principle to recurrent memory, we provide a control-theoretic account of what hidden states compute in continuous control and a mechanism for shaping them toward optimality.

Chemotherapy dose optimization can be formulated as a dynamic treatment regime, requiring sequential decisions under uncertainty that must balance tumor suppression against toxicity. However, most reinforcement learning approaches assume full observability of the patient state, a condition rarely met in clinical practice. We investigate whether memory-augmented policies can improve chemotherapy control under partial observability. To this end, we employ a recurrent TD3-based approach with separate LSTM actor-critic networks and evaluate it on the AhnChemoEnv benchmark from DTR-Bench, considering both off-policy and on-policy recurrent architectures against feed-forward TD3 and Soft Actor-Critic. Pharmacokinetic and pharmacodynamic variability are held fixed to isolate hidden-state uncertainty and observation noise and to avoid confounding effects from inter-patient variability. Across ten random seeds, recurrence yields modest benefit under full observability but substantially stronger and more stable performance under partial observability, with more consistent tumor suppression and improved normal-cell preservation. These findings indicate that memory-based policies are particularly beneficial when clinically relevant state information is incomplete or noisy.

In this paper, we propose a spatial-temporal learning-based distributed routing framework for dynamic Low Earth Orbit (LEO) satellite networks, where graph attention networks (GAT) and long short-term memory (LSTM) are integrated within a deep Q-network (DQN)-based architecture to enable distributed and adaptive routing decisions based on local observations. The routing problem is formulated as a partially observable Markov decision process (POMDP) to address partial observability under dynamic topology and time-varying traffic. Simulation results show that the proposed method significantly outperforms conventional and learning-based routing schemes in terms of throughput, packet loss, queue length, and end-to-end delay, while achieving proactive congestion avoidance with up to 23.26% queue reduction. In addition, the proposed approach maintains low computational overhead with negligible carbon emissions, demonstrating its efficiency from a Green AI perspective.

Closed-loop management of geological CO2 storage requires control policies that adapt to uncertain reservoir behavior while relying on observations that are realistically available during operation. This work formulates CO2 injection and brine-production control as a partially observable sequential decision problem and studies deployable deep reinforcement-learning controllers trained with high-fidelity reservoir simulation. We first compare privileged-state, well-only, history-conditioned, masking-curriculum, and asymmetric teacher-student model-free policies in order to quantify the value of temporal well-response information and training-time privileged simulator states. We then evaluate a latent model-based adaptation pipeline that reuses nominal latent dynamics and retunes controllers under known injector failure, leakage-induced dynamics and reward shift, and compartmentalized reservoir connectivity. The results show that history-conditioned policies recover nearly all of the privileged-state performance while using only deployable well-level information, and that latent model-based retuning outperforms direct model-free retuning under the same scenario-specific real-simulator budget in the abnormal operating cases. The proposed framework therefore provides a simulator-budget-aware alternative to repeated online history matching and re-optimization for closed-loop CO2 storage control.

In continual reinforcement learning (CRL), good performance requires never-ending learning, acting, and exploration in a big, partially observable world. Most CRL experiments have focused on loss of plasticity -- the inability to keep learning -- in one-off experiments where some unobservable non-stationarity is added to classic fully observable MDPs. Further, these experiments rarely consider the role of partial observability and the importance of CRL agents that use memory or recurrence. One potential reason for this focus on mitigating loss of plasticity without considering partial observability is that many partially-observable CRL environments are prohibitively expensive. In this paper, we introduce Forager, a light-weight partially-observable CRL environment with a constant memory footprint. We provide a set of experiments and sample tasks demonstrating that Forager is challenging for current CRL agents and yet also allows for in-depth study of those agents. We demonstrate that agents exhibit loss of plasticity, proposed mitigations can help, but that most useful is to leverage state construction. We conclude with a variant of Forager that generates an unending stream of new tasks to learn that clearly highlights the limitations of current CRL agents.

Physical awareness, especially in a large and dynamic environment, is shaped by sensing decisions that determine observability across space, time, and scale, while observations impact the quality of sensing decisions. This loopy information structure makes physical awareness a fundamentally challenging decision problem with partial observations. While in the past decade we have witnessed the unprecedented success of reinforcement learning (RL) in problems with full observability, decision problems with partial observation, such as POMDPs, remain largely open: real-world explorations are excessively costly, while sim-to-real pipeline suffer from unobserved viewpoints. We introduce AW4RE (Active World-model with 4D-informed Retrieval for Exploration), an awareness-centric generative world model that provides a sensor-native surrogate environment for exploring sensing queries. Conditioned on a queried sensing action, AW4RE estimates the action-conditioned observation process. This is done by combining 4D-informed evidence retrieval, action-conditioned geometric support with temporal coherence, and conditional generative completion. Experiments demonstrate that AW4RE produces more grounded and consistent predictions than geometry-aware generative baselines under extreme viewpoint shifts, temporal gaps, and sparse geometric support.

Penetration testing - the practice of simulating cyberattacks to identify vulnerabilities - is a complex sequential decision-making task that is inherently partially observable and features large action spaces. Existing RL simulators for this domain are CPU-bound and fixed to narrow scenarios, making it infeasible to train policies that generalize across networks. We present NASimJax, a JAX-native framework that formulates penetration testing as a Contextual POMDP and introduces a network generation pipeline producing structurally diverse, guaranteed-solvable scenarios. The framework reaches up to 80$\times$ higher environment throughput than previous simulators, enabling experiments on larger networks and tractable hyperparameter searches. We provide PPO and PQN baselines and conduct the first systematic evaluation of unsupervised environment design for penetration testing. We find that Prioritized Level Replay and ACCEL handle dense training distributions better than Domain Randomization, and that training on sparser topologies yields an implicit curriculum that improves generalization - even to topologies denser than those seen during training. A recurrent PPO variant confirms that these distributional findings are not an artifact of feed-forward policies. The code is available at: https://github.com/raphsimon/NASimJax.

Stabilizing a dynamical system is a fundamental problem that serves as a cornerstone for many complex tasks in the field of control systems. The problem becomes challenging when the system model is unknown. Among the Reinforcement Learning (RL) algorithms that have been successfully applied to solve problems pertaining to unknown linear dynamical systems, the policy gradient (PG) method stands out due to its ease of implementation and can solve the problem in a model-free manner. However, most of the existing works on PG methods for unknown linear dynamical systems assume full-state feedback. In this paper, we take a step towards model-free learning for partially observed linear dynamical systems with output feedback and focus on the fundamental stabilization problem of the system. We propose an algorithmic framework that stretches the boundary of PG methods to the problem without global convergence guarantees. We show that by leveraging zeroth-order PG update based on system trajectories and its convergence to stationary points, the proposed algorithms return a stabilizing output feedback policy for discrete-time linear dynamical systems. We also explicitly characterize the sample complexity of our algorithm and verify the effectiveness of the algorithm using numerical examples.

Penetration testing, the simulation of cyberattacks to identify security vulnerabilities, presents a sequential decision-making problem well-suited for reinforcement learning (RL) automation. Like many applications of RL to real-world problems, partial observability presents a major challenge, as it invalidates the Markov property present in Markov Decision Processes (MDPs). Partially Observable MDPs require history aggregation or belief state estimation to learn successful policies. We investigate stochastic, partially observable penetration testing scenarios over host networks of varying size, aiming to better reflect real-world complexity through more challenging and representative benchmarks. This approach leads to the development of more robust and transferable policies, which are crucial for ensuring reliable performance across diverse and unpredictable real-world environments. Using vanilla Proximal Policy Optimization (PPO) as a baseline, we compare a selection of PPO-based variants designed to mitigate partial observability, including frame-stacking, augmenting observations with historical information, and employing LSTM or TrXL architectures. We conduct a systematic empirical analysis of these algorithms across different host network sizes. We find that this task greatly benefits from history aggregation. Converging up to four times faster than other approaches. Manual inspection of the learned policies by the algorithms reveals clear distinctions and provides insights that go beyond quantitative results.

Reinforcement learning (RL) provides a principled framework for decision-making in partially observable environments, which can be modeled as Markov decision processes and compactly represented through dynamic decision Bayesian networks. Recent advances demonstrate that inference on sparse Bayesian networks can be accelerated using quantum rejection sampling combined with amplitude amplification, leading to a computational speedup in estimating acceptance probabilities. Building on this result, we introduce Quantum Bayesian Reinforcement Learning (QBRL), a hybrid quantum-classical look-ahead algorithm for model-based RL in partially observable environments. We present a rigorous, oracle-free time complexity analysis under fault-tolerant assumptions for the quantum device. Unlike standard treatments that assume a black-box oracle, we explicitly specify the inference process, allowing our bounds to more accurately reflect the true computational cost. We show that, for environments whose dynamics form a sparse Bayesian network, horizon-based near-optimal planning can be achieved sub-quadratically faster through quantum-enhanced belief updates. On the other hand, we show that there is no quantum speed-up for environments that are either fully observable, or characterized by Bayesian networks whose maximum in-degree is not small. Furthermore, we present numerical experiments benchmarking QBRL against its classical counterpart on simple yet illustrative decision-making tasks. Our results offer a detailed analysis of how the quantum computational advantage translates into decision-making performance, highlighting that the magnitude of the advantage can vary significantly across different deployment settings.

Recent advancements in multi-agent reinforcement learning (MARL) have opened up vast application prospects, such as swarm control of drones, collaborative manipulation by robotic arms, and multi-target encirclement. However, potential security threats during the MARL deployment need more attention and thorough investigation. Recent research reveals that attackers can rapidly exploit the victim's vulnerabilities, generating adversarial policies that result in the failure of specific tasks. For instance, reducing the winning rate of a superhuman-level Go AI to around 20%. Existing studies predominantly focus on two-player competitive environments, assuming attackers possess complete global state observation. In this study, we unveil, for the first time, the capability of attackers to generate adversarial policies even when restricted to partial observations of the victims in multi-agent competitive environments. Specifically, we propose a novel black-box attack (SUB-PLAY) that incorporates the concept of constructing multiple subgames to mitigate the impact of partial observability and suggests sharing transitions among subpolicies to improve attackers' exploitative ability. Extensive evaluations demonstrate the effectiveness of SUB-PLAY under three typical partial observability limitations. Visualization results indicate that adversarial policies induce significantly different activations of the victims' policy networks. Furthermore, we evaluate three potential defenses aimed at exploring ways to mitigate security threats posed by adversarial policies, providing constructive recommendations for deploying MARL in competitive environments.

Reinforcement learning (RL) performs well in continuous-control tasks, but evolving hidden disturbances create partial observability: the agent must infer decision-relevant latent dynamics from interaction history. This study investigates how observation history, action history, history length, and network structure affect recurrent Twin Delayed Deep Deterministic Policy Gradient (TD3) agents. Three recurrent architectures are evaluated under controlled disturbances with different temporal characteristics. Results show that action history is particularly important when observed responses depend on previous actions, and that processing past and current action-observation information within a unified temporal sequence improves performance compared with using separate branches. We also introduce H-TD3, which reuses recurrent states generated by the actor to initialize the critic, reducing duplicated sequence processing. The architectures are further tested in a simulation-based differential-drive rover motion-regulation task under hidden asymmetric wheel slip. Recurrent architectures retain their advantage under the physically motivated multiplicative wheel-slip model, while policies trained with abstract temporally structured disturbances transfer more effectively to previously unseen wheel-slip dynamics than policies trained without disturbances. These findings provide practical guidance for recurrent RL under partial observability and evolving hidden disturbances.

We study Nash equilibrium learning in partially observable Markov games (POMGs), a multi-agent reinforcement learning framework in which agents cannot fully observe the underlying state. Prior work in this setting relies on centralization or information sharing, and suffers from sample and computational complexity that scales exponentially in the number of players. We focus on a subclass of POMGs with independent state transitions, where agents remain coupled through their rewards, and assume that the underlying fully observed Markov game is a Markov potential game. For this class, we present an independent learning algorithm in which players, observing only their own actions and observations and without communication, jointly converge to an approximate Nash equilibrium. Due to partial observability, optimal policies may in general depend on the full action-observation history. Under a filter stability assumption, we show that policies based on finite history windows provide sufficient approximation guarantees. This enables us to approximate the POMG by a surrogate Markov game that is near-potential, leading to quasi-polynomial sample and computational complexity for independent Nash equilibrium learning in the underlying POMG.