Context-Adapted Multi-policy Ensemble Method for Generalization in Reinforcement Learning

Reinforcement learning (RL) has garnered significant attention for developing decision-making agents that aim to maximize rewards, specified by an external supervisor, within fully observable environments. However, many real-world problems involve partial or noisy observations, where agents cannot access complete and accurate information about the environment. These problems are commonly formulated as partially observable Markov decision processes (POMDPs). Previous studies have tackled RL in POMDPs by either incorporating the memory of past actions and observations or by inferring the true state of the environment from observed data. Nevertheless, aggregating observations and actions over time becomes impractical in problems with large decision-making time horizons and high-dimensional spaces. Furthermore, inference-based RL approaches often require many environmental samples to perform well, as they focus solely on reward maximization and neglect uncertainty in the inferred state. Active inference (AIF) is a framework naturally formulated in POMDPs and directs agents to select actions by minimizing a function called expected free energy (EFE). This supplies reward-maximizing (or exploitative) behavior, as in RL, with information-seeking (or exploratory) behavior. Despite this exploratory behavior of AIF, its use is limited to problems with small time horizons and discrete spaces due to the computational challenges associated with EFE. In this article, we propose a unified principle that establishes a theoretical connection between AIF and RL, enabling seamless integration of these two approaches and overcoming their limitations in continuous space POMDP settings. We substantiate our findings with rigorous theoretical analysis, providing novel perspectives for using AIF in designing and implementing artificial agents. Experimental results demonstrate the superior learning capabilities of our method compared to other alternative RL approaches in solving partially observable tasks with continuous spaces. Notably, our approach harnesses information-seeking exploration, enabling it to effectively solve reward-free problems and rendering explicit task reward design by an external supervisor optional.

Resolving structural uncertainty in natural resources management using POMDP approaches

Aiming to reduce energy consumption, minimize transmission latency, and optimize the long-term performance of Wireless Sensor Networks (WSN), this paper proposes an improved routing model of Directed Diffusion. We formulated the routing issue as a partially observable Markov decision process (POMDP), such that the optimal nodes selection policy yielded by maximizing the expected reward. Since the solution of POMDP is ready-made, the attention is concentrated only on the formulation of the routing issue into a POMDP model, more specifically, the definition of the tuple which contains states, actions, observations, transition probabilities, observation probabilities, and reward function. Our experimental study shows that the POMDP-based Directed Diffusion routing outperforms other routing techniques in terms of three widely used metrics including average dissipated energy, average delay, and average lifetime.

Deep reinforcement learning (RL) methods have achieved remarkable performance on challenging control tasks. Observations of the resulting behavior give the impression that the agent has constructed a generalized representation that supports insightful action decisions. We re‐examine what is meant by generalization in RL, and propose several definitions based on an agent's performance in on‐policy, off‐policy, and unreachable states. We propose a set of practical methods for evaluating agents with these definitions of generalization. We demonstrate these techniques on a common benchmark task for deep RL, and we show that the learned networks make poor decisions for states that differ only slightly from on‐policy states, even though those states are not selected adversarially. We focus our analyses on the deep Q‐networks (DQNs) that kicked off the modern era of deep RL. Taken together, these results call into question the extent to which DQNs learn generalized representations, and suggest that more experimentation and analysis is necessary before claims of representation learning can be supported.

During the last decades, collaborative robots capable of operating out of their cages are widely used in industry to assist humans in mundane and harsh manufacturing tasks. Although such robots are inherently safe by design, they are commonly accompanied by external sensors and other cyber-physical systems, to facilitate close cooperation with humans, which frequently render the collaborative ecosystem unsafe and prone to hazards. We introduce a method that capitalizes on partially observable Markov decision processes (POMDP) to amalgamate nominal actions of the system along with unsafe control actions posed by the System Theoretic Process Analysis (STPA). A decision-making mechanism that constantly prompts the system into a safer state is realized by providing situation awareness about the safety levels of the collaborative ecosystem by associating the system safety awareness with specific groups of selected actions. POMDP compensates the partial observability and uncertainty of the current state of the collaborative environment and creates safety screening policies that tend to make decisions that balance the system from unsafe to safe states in real time during the operational phase. The theoretical framework is assessed on a simulated human–robot collaborative scenario and proved capable of identifying loss and success scenarios.

Offline reinforcement learning (RL) is valuable in settings where online interactions with an environment are impractical. While such settings are often partially-observable, existing offline RL methods typically focus on fully-observable Markov decision processes (MDPs) rather than partially-observable MDPs (POMDPs). To help close that gap, we present an offline RL algorithm for POMDPs that leverages expert policies from simpler, fully-observable versions of environments in an asymmetric learning setting. We provide theoretical grounding for how overlap between MDPs and POMDPs can be exploited to improve learning in the partially-observable setting, and our experiments empirically demonstrate that our method significantly improves performance compared to existing state-of-the-art MDP offline RL algorithms.

Reinforcement learning (RL) offers a powerful paradigm for adaptive motion planning in complex driving environments. However, applying RL to autonomous driving remains challenging due to uncertainty from partial observability and the stochastic, multimodal behaviors of traffic participants. This paper presents a novel risk-aware RL framework for non-conservative motion planning under uncertainty. By integrating Partially Observable Markov Decision Processes (POMDP) with a deep RL-based policy optimization scheme, the proposed approach explicitly models aleatoric uncertainty via a Gaussian Mixture Bayesian Belief Updater and a time-varying risk field. Additionally, an Adaptive Context-aware Attention (ACA) module is employed to prioritize critical targets for enhanced interaction modeling dynamically. Extensive experiments on the CARLA simulator show that the framework generalizes well across diverse traffic conditions, improving average reward by 65.74% and 64.02% in low-speed dense and high-speed sparse scenarios. It remains robust in challenging situations such as overtaking and sudden lane changes in the PeMS dataset. Furthermore, distributed deployment tests confirm a real-time performance of 10 Hz on a hardware-in-the-loop platform, demonstrating the feasibility of practical deployment.

As the hardware and software complexity grows, it is unlikely for the power management hardware/software to have a full observation of the entire system status. In this paper, we propose a new modeling and optimization technique based on partially observable Markov decision process (POMDP) for robust power management, which can achieve near-optimal power savings, even when only partial system information is available. Three scenarios of partial observations that may occur in an embedded system are discussed and their modeling techniques are presented. The experimental results show that, compared with power management policy derived from traditional Markov decision process model that assumes the system is fully observable, the new power management technique gives significantly better performance and energy tradeoff.

Attitude control of a novel regional truss-braced wing (TBW) aircraft with low stability characteristics is addressed in this paper using Reinforcement Learning (RL). In recent years, RL has been increasingly employed in challenging applications, particularly, autonomous flight control. However, a significant predicament confronting discrete RL algorithms is the dimension limitation of the state-action table and difficulties in defining the elements of the RL environment. To address these issues, in this paper, a detailed mathematical model of the mentioned aircraft is first developed to shape an RL environment. Subsequently, Q-learning, the most prevalent discrete RL algorithm, will be implemented in both the Markov Decision Process (MDP) and Partially Observable Markov Decision Process (POMDP) frameworks to control the longitudinal mode of the proposed aircraft. In order to eliminate residual fluctuations that are a consequence of discrete action selection, and simultaneously track variable pitch angles, a Fuzzy Action Assignment (FAA) method is proposed to generate continuous control commands using the trained optimal Q-table. Accordingly, it will be proved that by defining a comprehensive reward function based on dynamic behavior considerations, along with observing all crucial states (equivalent to satisfying the Markov Property), the air vehicle would be capable of tracking the desired attitude in the presence of different uncertain dynamics including measurement noises, atmospheric disturbances, actuator faults, and model uncertainties where the performance of the introduced control system surpasses a well-tuned Proportional–Integral–Derivative (PID) controller.

Many robotic tasks, such as mobile manipulation, often require interaction with unstructured environments and are subject to imperfect sensing and actuation. This brings substantial uncertainty into the problems. Reasoningunder this uncertainty can provide higher level of robustness but is computationally significantly more challenging. More specifically, sequential decision making under motion and sensing uncertainty can be formulated in a principled form as a Partially Observable Markov Decision Process (POMDP). Solving POMDPs exactly is computationally intractable due to their exponential complexity with the number of states and the depth of the planning horizon, so called, the curse of dimensionality and the curse ofhistory, respectively. In this work, we propose a novel search-based robust planning framework that sample-efficiently finds a solution with theoretical suboptimalitybounds by leveraging multiple heuristics that are designed using domain knowledge. The main contributions of this work can be summarized as follows: 1) it works with generative models, in the absence of mathematical models, similarly to the reinforcement learning paradigm, but 2) it still exhibits high sample-efficiency by bootstrapping with the domain knowledge in the form of heuristics, and 3) it is empowered with effective guidance of the search toward the goal through the systematic employment of multipleheuristics. Its solution can be returned in an anytime fashion, and converges to the bounded suboptimality with theoretical guarantees. We validate the proposed frameworks through simulation and robot experiments,which includes 2D rover navigation, PR2 parts assembly, and full size mobile manipulator truck unloading tasks. The truck unloading task is especially interesting, where the custom-built robot is to unload up to several thousands of boxes from a truck. It is a challenging real-world manipulation problem in continuous space with excessively high uncertainty, and we demonstrate the efficacy of our approach in such a domain.

Foraging mammals exhibit a familiar yet poorly characterized phenomenon, ‘alternation’, a pause to sniff in the air preceded by the animal rearing on its hind legs or raising its head. Rodents spontaneously alternate in the presence of airflow, suggesting that alternation serves an important role during plume-tracking. To test this hypothesis, we combine fully resolved simulations of turbulent odor transport and Bellman optimization methods for decision-making under partial observability. We show that an agent trained to minimize search time in a realistic odor plume exhibits extensive alternation together with the characteristic cast-and-surge behavior observed in insects. Alternation is linked with casting and occurs more frequently far downwind of the source, where the likelihood of detecting airborne cues is higher relative to ground cues. Casting and alternation emerge as complementary tools for effective exploration with sparse cues. A model based on marginal value theory captures the interplay between casting, surging, and alternation.

This paper proposes a novel direct policy search (DPS) method with model selection for partially observed Markov decision processes (POMDPs). DPSs have been standard for learning POMDPs due to their computational efficiency and natural ability to maximize total rewards. An important open challenge for the best use of DPS methods is model selection, i.e., determination of the proper dimensionality of hidden states and complexity of policy functions, to mitigate overfitting in highly-flexible model representations of POMDPs. This paper bridges Bayesian inference and reward maximization and derives marginalized weighted log-likelihood~(MWL) for POMDPs which takes both advantages of Bayesian model selection and DPS. Then we propose factorized asymptotic Bayesian policy search (FABPS) to explore the model and the policy which maximizes MWL by expanding recently-developed factorized asymptotic Bayesian inference. Experimental results show that FABPS outperforms state-of-the-art model selection methods for POMDPs, with respect both to model selection and to expected total rewards.

In this article, we consider a subclass of partially observable Markov decision process (POMDP) problems which we termed confounding POMDPs. In these types of POMDPs, temporal difference (TD)-based reinforcement learning (RL) algorithms struggle, as TD error cannot be easily derived from observations. We solve these types of problems using a new bio-inspired neural architecture that combines a modulated Hebbian network (MOHN) with deep Q-network (DQN), which we call modulated Hebbian plus Q-network architecture (MOHQA). The key idea is to use a Hebbian network with rarely correlated bio-inspired neural traces to bridge temporal delays between actions and rewards when confounding observations and sparse rewards result in inaccurate TD errors. In MOHQA, DQN learns low-level features and control, while the MOHN contributes to high-level decisions by associating rewards with past states and actions. Thus, the proposed architecture combines two modules with significantly different learning algorithms, a Hebbian associative network and a classical DQN pipeline, exploiting the advantages of both. Simulations on a set of POMDPs and on the Malmo environment show that the proposed algorithm improved DQN's results and even outperformed control tests with advantage-actor critic (A2C), quantile regression DQN with long short-term memory (QRDQN + LSTM), Monte Carlo policy gradient (REINFORCE), and aggregated memory for reinforcement learning (AMRL) algorithms on most difficult POMDPs with confounding stimuli and sparse rewards.

Ambiguous partially observable Markov decision processes: Structural results and applications

In this article, we study the problem of controlling a partially observed Markov decision process (POMDP) to either aid or hinder the estimation of its state trajectory. We encode the estimation objectives via the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">smoother entropy</i> , which is the conditional entropy of the state trajectory given measurements and controls. Consideration of the smoother entropy contrasts with previous approaches that instead resort to marginal (or instantaneous) state entropies due to tractability concerns. By establishing novel expressions for the smoother entropy in terms of the POMDP belief state, we show that both the problems of minimizing and maximizing the smoother entropy in POMDPs can surprisingly be reformulated as belief-state Markov decision processes with concave cost and value functions. The significance of these reformulations is that they render the smoother entropy a tractable optimization objective, with structural properties amenable to the use of standard POMDP solution techniques for <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">both</i> active estimation and obfuscation. Simulations illustrate that optimization of the smoother entropy leads to superior trajectory estimation and obfuscation compared to alternative approaches.