Safe reinforcement learning (RL) focuses on training reward-maximizing agents subject to pre-defined safety constraints. Yet, learning versatile safe policies that can adapt to varying safety constraint requirements during deployment without retraining remains a largely unexplored and challenging area. In this work, we formulate the versatile safe RL problem and consider two primary requirements: training efficiency and zero-shot adaptation capability. To address them, we introduce the Conditioned Constrained Policy Optimization (CCPO) framework, consisting of two key modules: (1) Versatile Value Estimation (VVE) for approximating value functions under unseen threshold conditions, and (2) Conditioned Variational Inference (CVI) for encoding arbitrary constraint thresholds during policy optimization. Our extensive experiments demonstrate that CCPO outperforms the baselines in terms of safety and task performance while preserving zero-shot adaptation capabilities to different constraint thresholds data-efficiently. This makes our approach suitable for real-world dynamic applications.

Safety assurance of Reinforcement Learning (RL) is critical for exploration in real-world scenarios. In handling the Constrained Markov Decision Process, current approaches experience intrinsic difficulties in trading-off between optimality and feasibility. Direct optimization methods cannot strictly guarantee state-wise in-training safety while projection-based methods are usually inefficient and correct actions through lengthy iterations. To address these two challenges, this paper proposes an adaptive surrogate chance constraint for the safety cost, and a hierarchical architecture that corrects actions produced by the upper policy layer via a fast Quasi-Newton method. Theoretical analysis indicates that the relaxed probabilistic constraint can sufficiently guarantee forward invariance to the safe set. We validate the proposed method on 4 simulated and real-world safety-critical robotic tasks. Results indicate that the proposed method can efficiently enforce safety (nearly zero-violation), while preserving optimality (+23.8%), robustness and generalizability to stochastic real-world settings.

Self-play (SP) is a popular multi-agent reinforcement learning (MARL) framework for solving competitive games, where each agent optimizes policy by treating others as part of the environment. Despite the empirical successes, the theoretical properties of SP-based methods are limited to two-player zero-sum games. However, for mixed cooperative-competitive games where agents on the same team need to cooperate with each other, we can show a simple counter-example where SP-based methods cannot converge to a global Nash equilibrium (NE) with high probability. Alternatively, Policy-Space Response Oracles (PSRO) is an iterative framework for learning NE, where the best responses w.r.t. previous policies are learned in each iteration. PSRO can be directly extended to mixed cooperative-competitive settings by jointly learning team best responses with all convergence properties unchanged. However, PSRO requires repeatedly training joint policies from scratch till convergence, which makes it hard to scale to complex games. In this work, we develop a novel algorithm, Fictitious Cross-Play (FXP), which inherits the benefits from both frameworks. FXP simultaneously trains an SP-based main policy and a counter population of best response policies. The main policy is trained by fictitious self-play and cross-play against the counter population, while the counter policies are trained as the best responses to the main policy's past versions. We validate our method in matrix games and show that FXP converges to global NEs while SP methods fail. We also conduct experiments in a gridworld domain, where FXP achieves higher Elo ratings and lower exploitabilities than baselines, and a more challenging football game, where FXP defeats SOTA models with over 94% win rate.

Multi-Agent Reinforcement Learning (MARL) has become a promising solution for constructing a multi-agent autonomous driving system (MADS) in complex and dense scenarios. But most methods consider agents acting selfishly, which leads to conflict behaviors. Some existing works incorporate the concept of social value orientation (SVO) to promote coordination, but they lack the knowledge of other agents' SVOs, resulting in conservative maneuvers. In this paper, we aim to tackle the mentioned problem by enabling the agents to understand other agents' SVOs. To accomplish this, we propose a two-stage system framework. Firstly, we train a policy by allowing the agents to share their ground truth SVOs to establish a coordinated traffic flow. Secondly, we develop a recognition network that estimates agents' SVOs and integrates it with the policy trained in the first stage. Experiments demonstrate that our developed method significantly improves the performance of the driving policy in MADS compared to two state-of-the-art MARL algorithms.

Recently, it has been shown that for offline deep reinforcement learning (DRL), pre-training Decision Transformer with a large language corpus can improve downstream performance (Reid et al., 2022). A natural question to ask is whether this performance gain can only be achieved with language pre-training, or can be achieved with simpler pre-training schemes which do not involve language. In this paper, we first show that language is not essential for improved performance, and indeed pre-training with synthetic IID data for a small number of updates can match the performance gains from pre-training with a large language corpus; moreover, pre-training with data generated by a one-step Markov chain can further improve the performance. Inspired by these experimental results, we then consider pre-training Conservative Q-Learning (CQL), a popular offline DRL algorithm, which is Q-learning-based and typically employs a Multi-Layer Perceptron (MLP) backbone. Surprisingly, pre-training with simple synthetic data for a small number of updates can also improve CQL, providing consistent performance improvement on D4RL Gym locomotion datasets. The results of this paper not only illustrate the importance of pre-training for offline DRL but also show that the pre-training data can be synthetic and generated with remarkably simple mechanisms. It is well-known that pre-training can provide significant boosts in performance and robustness for downstream tasks, both for Natural Language Processing (NLP) and Computer Vision (CV). Recently, in the field of Deep Reinforcement Learning (DRL), research on pre-training is also becoming increasingly popular. An important step in the direction of pre-training DRL models is the recent paper by Reid et al. (2022), which showed that for Decision Transformer (Chen et al., 2021), pretraining with the Wikipedia corpus can significantly improve the performance of the downstream offline RL task. Reid et al. (2022) further showed that pre-training on predicting pixel sequences can hurt performance. The authors state that their results indicate "a foreseeable future where everyone should use a pre-trained language model for offline RL".

In this paper we provide a framework to cope with two problems: (i) the fragility of reinforcement learning due to modeling uncertainties because of the mismatch between controlled laboratory/simulation and real-world conditions and (ii) the prohibitive computational cost of stochastic optimal control. We approach both problems by using reinforcement learning to solve the stochastic dynamic programming equation. The resulting reinforcement learning controller is safe with respect to several types of constraints and it can actively learn about the modeling uncertainties. Unlike exploration and exploitation, probing and safety are employed automatically by the controller itself, resulting real-time learning. A simulation example demonstrates the efficacy of the proposed approach.

Reinforcement learning (RL) has shown promising results for real-time control systems, including the domain of plasma magnetic control. However, there are still significant drawbacks compared to traditional feedback control approaches for magnetic confinement. In this work, we address key drawbacks of the RL method; achieving higher control accuracy for desired plasma properties, reducing the steady-state error, and decreasing the required time to learn new tasks. We build on top of \cite{degrave2022magnetic}, and present algorithmic improvements to the agent architecture and training procedure. We present simulation results that show up to 65\% improvement in shape accuracy, achieve substantial reduction in the long-term bias of the plasma current, and additionally reduce the training time required to learn new tasks by a factor of 3 or more. We present new experiments using the upgraded RL-based controllers on the TCV tokamak, which validate the simulation results achieved, and point the way towards routinely achieving accurate discharges using the RL approach.

Real-world deployment of machine learning models is challenging because data evolves over time. While no model can work when data evolves in an arbitrary fashion, if there is some pattern to these changes, we might be able to design methods to address it. This paper addresses situations when data evolves gradually. We introduce a time-varying propensity score that can detect gradual shifts in the distribution of data which allows us to selectively sample past data to update the model -- not just similar data from the past like that of a standard propensity score but also data that evolved in a similar fashion in the past. The time-varying propensity score is quite general: we demonstrate different ways of implementing it and evaluate it on a variety of problems ranging from supervised learning (e.g., image classification problems) where data undergoes a sequence of gradual shifts, to reinforcement learning tasks (e.g., robotic manipulation and continuous control) where data shifts as the policy or the task changes.

We present the problem of reinforcement learning with exogenous termination. We define the Termination Markov Decision Process (TerMDP), an extension of the MDP framework, in which episodes may be interrupted by an external non-Markovian observer. This formulation accounts for numerous real-world situations, such as a human interrupting an autonomous driving agent for reasons of discomfort. We learn the parameters of the TerMDP and leverage the structure of the estimation problem to provide state-wise confidence bounds. We use these to construct a provably-efficient algorithm, which accounts for termination, and bound its regret. Motivated by our theoretical analysis, we design and implement a scalable approach, which combines optimism (w.r.t. termination) and a dynamic discount factor, incorporating the termination probability. We deploy our method on high-dimensional driving and MinAtar benchmarks. Additionally, we test our approach on human data in a driving setting. Our results demonstrate fast convergence and significant improvement over various baseline approaches.

GFlowNets are probabilistic models that learn a stochastic policy that sequentially generates compositional structures, such as molecular graphs. They are trained with the objective of sampling such objects with probability proportional to the object's reward. Among GFlowNets, the temperature-conditional GFlowNets represent a family of policies indexed by temperature, and each is associated with the correspondingly tempered reward function. The major benefit of temperature-conditional GFlowNets is the controllability of GFlowNets' exploration and exploitation through adjusting temperature. We propose Learning to Scale Logits for temperature-conditional GFlowNets (LSL-GFN), a novel architectural design that greatly accelerates the training of temperature-conditional GFlowNets. It is based on the idea that previously proposed temperature-conditioning approaches introduced numerical challenges in the training of the deep network because different temperatures may give rise to very different gradient profiles and ideal scales of the policy's logits. We find that the challenge is greatly reduced if a learned function of the temperature is used to scale the policy's logits directly. We empirically show that our strategy dramatically improves the performances of GFlowNets, outperforming other baselines, including reinforcement learning and sampling methods, in terms of discovering diverse modes in multiple biochemical tasks.