Motivated by the increasing deployment of reinforcement learning in the real world, involving a large consumption of personal data, we introduce a differentially private (DP) policy gradient algorithm. We show that, in this setting, the introduction of Differential Privacy can be reduced to the computation of appropriate trust regions, thus avoiding the sacrifice of theoretical properties of the DP-less methods. Therefore, we show that it is possible to find the right trade-off between privacy noise and trust-region size to obtain a performant differentially private policy gradient algorithm. We then outline its performance empirically on various benchmarks. Our results and the complexity of the tasks addressed represent a significant improvement over existing DP algorithms in online RL.

Reinforcement learning (RL) for continuous control often requires large amounts of online interaction data. Value-based RL methods can mitigate this burden by offering relatively high sample efficiency. Some studies further enhance sample efficiency by incorporating offline demonstration data to "kick-start" training, achieving promising results in continuous control. However, they typically compute the Q-function independently for each action dimension, neglecting interdependencies and making it harder to identify optimal actions when learning from suboptimal data, such as non-expert demonstration and online-collected data during the training process. To address these issues, we propose Auto-Regressive Soft Q-learning (ARSQ), a value-based RL algorithm that models Q-values in a coarse-to-fine, auto-regressive manner. First, ARSQ decomposes the continuous action space into discrete spaces in a coarse-to-fine hierarchy, enhancing sample efficiency for fine-grained continuous control tasks. Next, it auto-regressively predicts dimensional action advantages within each decision step, enabling more effective decision-making in continuous control tasks. We evaluate ARSQ on two continuous control benchmarks, RLBench and D4RL, integrating demonstration data into online training. On D4RL, which includes non-expert demonstrations, ARSQ achieves an average $1.62\times$ performance improvement over SOTA value-based baseline. On RLBench, which incorporates expert demonstrations, ARSQ surpasses various baselines, demonstrating its effectiveness in learning from suboptimal online-collected data.

$Q$-learning is one of the most fundamental reinforcement learning algorithms. Previously, it is widely believed that $Q$-learning with linear function approximation (i.e., linear $Q$-learning) suffers from possible divergence. This paper instead establishes the first $L^2$ convergence rate of linear $Q$-learning to a bounded set. Notably, we do not make any modification to the original linear $Q$-learning algorithm, do not make any Bellman completeness assumption, and do not make any near-optimality assumption on the behavior policy. All we need is an $\epsilon$-softmax behavior policy with an adaptive temperature. The key to our analysis is the general result of stochastic approximations under Markovian noise with fast-changing transition functions. As a side product, we also use this general result to establish the $L^2$ convergence rate of tabular $Q$-learning with an $\epsilon$-softmax behavior policy, for which we rely on a novel pseudo-contraction property of the weighted Bellman optimality operator.

In-Context Reinforcement Learning (ICRL) represents a promising paradigm for developing generalist agents that learn at inference time through trial-and-error interactions, analogous to how large language models adapt contextually, but with a focus on reward maximization. However, the scalability of ICRL beyond toy tasks and single-domain settings remains an open challenge. In this work, we present the first steps toward scaling ICRL by introducing a fixed, cross-domain model capable of learning behaviors through in-context reinforcement learning. Our results demonstrate that Algorithm Distillation, a framework designed to facilitate ICRL, offers a compelling and competitive alternative to expert distillation to construct versatile action models. These findings highlight the potential of ICRL as a scalable approach for generalist decision-making systems. Code to be released at https://github.com/dunnolab/vintix

In reinforcement learning for partially observable environments, many successful algorithms were developed within the asymmetric learning paradigm. This paradigm leverages additional state information available at training time for faster learning. Although the proposed learning objectives are usually theoretically sound, these methods still lack a theoretical justification for their potential benefits. We propose such a justification for asymmetric actor-critic algorithms with linear function approximators by adapting a finite-time convergence analysis to this setting. The resulting finite-time bound reveals that the asymmetric critic eliminates an error term arising from aliasing in the agent state.

Efficient job allocation in complex scheduling problems poses significant challenges in real-world applications. In this report, we propose a novel approach that leverages the power of Reinforcement Learning (RL) and Graph Neural Networks (GNNs) to tackle the Job Allocation Problem (JAP). The JAP involves allocating a maximum set of jobs to available resources while considering several constraints. Our approach enables learning of adaptive policies through trial-and-error interactions with the environment while exploiting the graph-structured data of the problem. By leveraging RL, we eliminate the need for manual annotation, a major bottleneck in supervised learning approaches. Experimental evaluations on synthetic and real-world data demonstrate the effectiveness and generalizability of our proposed approach, outperforming baseline algorithms and showcasing its potential for optimizing job allocation in complex scheduling problems.

We seek measurable properties of AI agents that make them better or worse teammates from the subjective perspective of human collaborators. Our experiments use the cooperative card game Hanabi -- a common benchmark for AI-teaming research. We first evaluate AI agents on a set of objective metrics based on task performance, information theory, and game theory, which are measurable without human interaction. Next, we evaluate subjective human preferences toward AI teammates in a large-scale (N=241) human-AI teaming experiment. Finally, we correlate the AI-only objective metrics with the human subjective preferences. Our results refute common assumptions from prior literature on reinforcement learning, revealing new correlations between AI behaviors and human preferences. We find that the final game score a human-AI team achieves is less predictive of human preferences than esoteric measures of AI action diversity, strategic dominance, and ability to team with other AI. In the future, these correlations may help shape reward functions for training human-collaborative AI.

The effectiveness of credit assignment in reinforcement learning (RL) when dealing with high-dimensional data is influenced by the success of representation learning via deep neural networks, and has implications for the sample efficiency of deep RL algorithms. Input decorrelation has been previously introduced as a method to speed up optimization in neural networks, and has proven impactful in both efficient deep learning and as a method for effective representation learning for deep RL algorithms. We propose a novel approach to online decorrelation in deep RL based on the decorrelated backpropagation algorithm that seamlessly integrates the decorrelation process into the RL training pipeline. Decorrelation matrices are added to each layer, which are updated using a separate decorrelation learning rule that minimizes the total decorrelation loss across all layers, in parallel to minimizing the usual RL loss. We used our approach in combination with the soft actor-critic (SAC) method, which we refer to as decorrelated soft actor-critic (DSAC). Experiments on the Atari 100k benchmark with DSAC shows, compared to the regular SAC baseline, faster training in five out of the seven games tested and improved reward performance in two games with around 50% reduction in wall-clock time, while maintaining performance levels on the other games. These results demonstrate the positive impact of network-wide decorrelation in deep RL for speeding up its sample efficiency through more effective credit assignment.

Symbolic task representation is a powerful tool for encoding human instructions and domain knowledge. Such instructions guide robots to accomplish diverse objectives and meet constraints through reinforcement learning (RL). Most existing methods are based on fixed mappings from environmental states to symbols. However, in inspection tasks, where equipment conditions must be evaluated from multiple perspectives to avoid errors of oversight, robots must fulfill the same symbol from different states. To help robots respond to flexible symbol mapping, we propose representing symbols and their mapping specifications separately within an RL policy. This approach imposes on RL policy to learn combinations of symbolic instructions and mapping specifications, requiring an efficient learning framework. To cope with this issue, we introduce an approach for learning flexible policies called Symbolic Instructions with Adjustable Mapping Specifications (SIAMS). This paper represents symbolic instructions using linear temporal logic (LTL), a formal language that can be easily integrated into RL. Our method addresses the diversified completion patterns of instructions by (1) a specification-aware state modulation, which embeds differences in mapping specifications in state features, and (2) a symbol-number-based task curriculum, which gradually provides tasks according to the learning's progress. Evaluations in 3D simulations with discrete and continuous action spaces demonstrate that our method outperforms context-aware multitask RL comparisons.

This article introduces a curriculum learning approach to develop a reinforcement learning-based robust stabilizing controller for a Quadrotor that meets predefined performance criteria. The learning objective is to achieve desired positions from random initial conditions while adhering to both transient and steady-state performance specifications. This objective is challenging for conventional one-stage end-to-end reinforcement learning, due to the strong coupling between position and orientation dynamics, the complexity in designing and tuning the reward function, and poor sample efficiency, which necessitates substantial computational resources and leads to extended convergence times. To address these challenges, this work decomposes the learning objective into a three-stage curriculum that incrementally increases task complexity. The curriculum begins with learning to achieve stable hovering from a fixed initial condition, followed by progressively introducing randomization in initial positions, orientations and velocities. A novel additive reward function is proposed, to incorporate transient and steady-state performance specifications. The results demonstrate that the Proximal Policy Optimization (PPO)-based curriculum learning approach, coupled with the proposed reward structure, achieves superior performance compared to a single-stage PPO-trained policy with the same reward function, while significantly reducing computational resource requirements and convergence time. The curriculum-trained policy's performance and robustness are thoroughly validated under random initial conditions and in the presence of disturbances.