Large language models (LLMs), despite their extensive pretraining on diverse datasets, require effective alignment to human preferences for practical and reliable deployment. Conventional alignment methods typically employ off-policy learning and depend on human-annotated datasets, which limits their broad applicability and introduces distribution shift issues during training. To address these challenges, we propose Self-Generated Preference Optimization based on Self-Improver (SGPO), an innovative alignment framework that leverages an on-policy self-improving mechanism. Specifically, the improver refines responses from a policy model to self-generate preference data for direct preference optimization (DPO) of the policy model. Here, the improver and policy are unified into a single model, and in order to generate higher-quality preference data, this self-improver learns to make incremental yet discernible improvements to the current responses by referencing supervised fine-tuning outputs. Experimental results on AlpacaEval 2.0 and Arena-Hard show that the proposed SGPO significantly improves performance over DPO and baseline self-improving methods without using external preference data.

Model-free reinforcement learning has emerged as a powerful method for developing robust robot control policies capable of navigating through complex and unstructured environments. The effectiveness of these methods hinges on two essential elements: (1) the use of massively parallel physics simulations to expedite policy training, and (2) an environment generator tasked with crafting sufficiently challenging yet attainable environments to facilitate continuous policy improvement. Existing methods of outdoor environment generation often rely on heuristics constrained by a set of parameters, limiting the diversity and realism. In this work, we introduce ADEPT, a novel \textbf{A}daptive \textbf{D}iffusion \textbf{E}nvironment for \textbf{P}olicy \textbf{T}ransfer in the zero-shot sim-to-real fashion that leverages Denoising Diffusion Probabilistic Models to dynamically expand existing training environments by adding more diverse and complex environments adaptive to the current policy. ADEPT guides the diffusion model's generation process through initial noise optimization, blending noise-corrupted environments from existing training environments weighted by the policy's performance in each corresponding environment. By manipulating the noise corruption level, ADEPT seamlessly transitions between generating similar environments for policy fine-tuning and novel ones to expand training diversity. To benchmark ADEPT in off-road navigation, we propose a fast and effective multi-layer map representation for wild environment generation. Our experiments show that the policy trained by ADEPT outperforms both procedural generated and natural environments, along with popular navigation methods.

The main contributions are: (a) a new algorithm, inspired by UCRL\gamma for the discounted setting (b) a theoretical analysis showing that the new algorithm enjoys sample-complexity guarantees that scale optimally with the horizon, the accuracy and the confidence. The result is sub-optimal only with respect to the size of the state/space, where it scales potentially quadratically while known lower bounds are linear in the state-space.

Experience replay serves as a key component in the success of online reinforcement learning (RL). Prioritized experience replay (PER) reweights experiences by the temporal difference (TD) error empirically enhancing the performance. However, few works have explored the motivation of using TD error. In this work, we provide an alternative perspective on TD-error-based reweighting. We show the connections between the experience prioritization and occupancy optimization. By using a regularized RL objective with $f-$divergence regularizer and employing its dual form, we show that an optimal solution to the objective is obtained by shifting the distribution of off-policy data in the replay buffer towards the on-policy optimal distribution using TD-error-based occupancy ratios. Our derivation results in a new pipeline of TD error prioritization. We specifically explore the KL divergence as the regularizer and obtain a new form of prioritization scheme, the regularized optimal experience replay (ROER). We evaluate the proposed prioritization scheme with the Soft Actor-Critic (SAC) algorithm in continuous control MuJoCo and DM Control benchmark tasks where our proposed scheme outperforms baselines in 6 out of 11 tasks while the results of the rest match with or do not deviate far from the baselines. Further, using pretraining, ROER achieves noticeable improvement on difficult Antmaze environment where baselines fail, showing applicability to offline-to-online fine-tuning. Code is available at \url{https://github.com/XavierChanglingLi/Regularized-Optimal-Experience-Replay}.

In Multi-agent Reinforcement Learning (MARL), accurately perceiving opponents' strategies is essential for both cooperative and adversarial contexts, particularly within dynamic environments. While Proximal Policy Optimization (PPO) and related algorithms such as Actor-Critic with Experience Replay (ACER), Trust Region Policy Optimization (TRPO), and Deep Deterministic Policy Gradient (DDPG) perform well in single-agent, stationary environments, they suffer from high variance in MARL due to non-stationary and hidden policies of opponents, leading to diminished reward performance. Additionally, existing methods in MARL face significant challenges, including the need for inter-agent communication, reliance on explicit reward information, high computational demands, and sampling inefficiencies. These issues render them less effective in continuous environments where opponents may abruptly change their policies without prior notice. Against this background, we present OPS-DeMo (Online Policy Switch-Detection Model), an online algorithm that employs dynamic error decay to detect changes in opponents' policies. OPS-DeMo continuously updates its beliefs using an Assumed Opponent Policy (AOP) Bank and selects corresponding responses from a pre-trained Response Policy Bank. Each response policy is trained against consistently strategizing opponents, reducing training uncertainty and enabling the effective use of algorithms like PPO in multi-agent environments. Comparative assessments show that our approach outperforms PPO-trained models in dynamic scenarios like the Predator-Prey setting, providing greater robustness to sudden policy shifts and enabling more informed decision-making through precise opponent policy insights.

Proximal Policy Optimization (PPO) is a popular model-free reinforcement learning algorithm, esteemed for its simplicity and efficacy. However, due to its inherent on-policy nature, its proficiency in harnessing data from disparate policies is constrained. This paper introduces a novel off-policy extension to the original PPO method, christened Transductive Off-policy PPO (ToPPO). Herein, we provide theoretical justification for incorporating off-policy data in PPO training and prudent guidelines for its safe application. Our contribution includes a novel formulation of the policy improvement lower bound for prospective policies derived from off-policy data, accompanied by a computationally efficient mechanism to optimize this bound, underpinned by assurances of monotonic improvement. Comprehensive experimental results across six representative tasks underscore ToPPO's promising performance.