We introduce a method to address goal misgeneralization in reinforcement learning (RL), leveraging Large Language Model (LLM) feedback during training. Goal misgeneralization, a type of robustness failure in RL occurs when an agent retains its capabilities out-of-distribution yet pursues a proxy rather than the intended one. Our approach utilizes LLMs to analyze an RL agent's policies during training and identify potential failure scenarios. The RL agent is then deployed in these scenarios, and a reward model is learnt through the LLM preferences and feedback. This LLM-informed reward model is used to further train the RL agent on the original dataset. We apply our method to a maze navigation task, and show marked improvements in goal generalization, especially in cases where true and proxy goals are somewhat distinguishable and behavioral biases are pronounced. This study demonstrates how the LLM, despite its lack of task proficiency, can efficiently supervise RL agents, providing scalable oversight and valuable insights for enhancing goal-directed learning in RL through the use of LLMs.

Command and control (C2) paths for issuing commands to malware are sometimes the only indicators of its existence within networks. Identifying potential C2 channels is often a manually driven process that involves a deep understanding of cyber tradecraft. Efforts to improve discovery of these channels through using a reinforcement learning (RL) based approach that learns to automatically carry out C2 attack campaigns on large networks, where multiple defense layers are in place serves to drive efficiency for network operators. In this paper, we model C2 traffic flow as a three-stage process and formulate it as a Markov decision process (MDP) with the objective to maximize the number of valuable hosts whose data is exfiltrated. The approach also specifically models payload and defense mechanisms such as firewalls which is a novel contribution. The attack paths learned by the RL agent can in turn help the blue team identify high-priority vulnerabilities and develop improved defense strategies. The method is evaluated on a large network with more than a thousand hosts and the results demonstrate that the agent can effectively learn attack paths while avoiding firewalls.

Recent advances in Multi-Agent Reinforcement Learning have prompted the modeling of intricate interactions between agents in simulated environments. In particular, the predator-prey dynamics have captured substantial interest and various simulations been tailored to unique requirements. To prevent further time-intensive developments, we introduce Aquarium, a comprehensive Multi-Agent Reinforcement Learning environment for predator-prey interaction, enabling the study of emergent behavior. Aquarium is open source and offers a seamless integration of the PettingZoo framework, allowing a quick start with proven algorithm implementations. It features physics-based agent movement on a two-dimensional, edge-wrapping plane. The agent-environment interaction (observations, actions, rewards) and the environment settings (agent speed, prey reproduction, predator starvation, and others) are fully customizable. Besides a resource-efficient visualization, Aquarium supports to record video files, providing a visual comprehension of agent behavior. To demonstrate the environment's capabilities, we conduct preliminary studies which use PPO to train multiple prey agents to evade a predator. In accordance to the literature, we find Individual Learning to result in worse performance than Parameter Sharing, which significantly improves coordination and sample-efficiency.

With recent advancements in quantum computing technology, optimizing quantum circuits and ensuring reliable quantum state preparation have become increasingly vital. Traditional methods often demand extensive expertise and manual calculations, posing challenges as quantum circuits grow in qubit- and gate-count. Therefore, harnessing machine learning techniques to handle the growing variety of gate-to-qubit combinations is a promising approach. In this work, we introduce a comprehensive reinforcement learning environment for quantum circuit synthesis, where circuits are constructed utilizing gates from the the Clifford+T gate set to prepare specific target states. Our experiments focus on exploring the relationship between the depth of synthesized quantum circuits and the circuit depths used for target initialization, as well as qubit count. We organize the environment configurations into multiple evaluation levels and include a range of well-known quantum states for benchmarking purposes. We also lay baselines for evaluating the environment using Proximal Policy Optimization. By applying the trained agents to benchmark tests, we demonstrated their ability to reliably design minimal quantum circuits for a selection of 2-qubit Bell states.

We address the challenge of learning safe and robust decision policies in presence of uncertainty in context of the real scientific problem of adaptive resource oversubscription to enhance resource efficiency while ensuring safety against resource congestion risk. Traditional supervised prediction or forecasting models are ineffective in learning adaptive policies whereas standard online optimization or reinforcement learning is difficult to deploy on real systems. Offline methods such as imitation learning (IL) are ideal since we can directly leverage historical resource usage telemetry. But, the underlying aleatoric uncertainty in such telemetry is a critical bottleneck. We solve this with our proposed novel chance-constrained imitation learning framework, which ensures implicit safety against uncertainty in a principled manner via a combination of stochastic (chance) constraints on resource congestion risk and ensemble value functions. This leads to substantial ($\approx 3-4\times$) improvement in resource efficiency and safety in many oversubscription scenarios, including resource management in cloud services.

Training recurrent neural networks typically relies on backpropagation through time (BPTT). BPTT depends on forward and backward passes to be completed, rendering the network locked to these computations before loss gradients are available. Recently, Jaderberg et al. proposed synthetic gradients to alleviate the need for full BPTT. In their implementation synthetic gradients are learned through a mixture of backpropagated gradients and bootstrapped synthetic gradients, analogous to the temporal difference (TD) algorithm in Reinforcement Learning (RL). However, as in TD learning, heavy use of bootstrapping can result in bias which leads to poor synthetic gradient estimates. Inspired by the accumulate $\mathrm{TD}(\lambda)$ in RL, we propose a fully online method for learning synthetic gradients which avoids the use of BPTT altogether: accumulate $BP(\lambda)$. As in accumulate $\mathrm{TD}(\lambda)$, we show analytically that accumulate $\mathrm{BP}(\lambda)$ can control the level of bias by using a mixture of temporal difference errors and recursively defined eligibility traces. We next demonstrate empirically that our model outperforms the original implementation for learning synthetic gradients in a variety of tasks, and is particularly suited for capturing longer timescales. Finally, building on recent work we reflect on accumulate $\mathrm{BP}(\lambda)$ as a principle for learning in biological circuits. In summary, inspired by RL principles we introduce an algorithm capable of bias-free online learning via synthetic gradients.

Quantum computing offers efficient encapsulation of high-dimensional states. In this work, we propose a novel quantum reinforcement learning approach that combines the Advantage Actor-Critic algorithm with variational quantum circuits by substituting parts of the classical components. This approach addresses reinforcement learning's scalability concerns while maintaining high performance. We empirically test multiple quantum Advantage Actor-Critic configurations with the well known Cart Pole environment to evaluate our approach in control tasks with continuous state spaces. Our results indicate that the hybrid strategy of using either a quantum actor or quantum critic with classical post-processing yields a substantial performance increase compared to pure classical and pure quantum variants with similar parameter counts. They further reveal the limits of current quantum approaches due to the hardware constraints of noisy intermediate-scale quantum computers, suggesting further research to scale hybrid approaches for larger and more complex control tasks.

Multi-Agent Reinforcement Learning is becoming increasingly more important in times of autonomous driving and other smart industrial applications. Simultaneously a promising new approach to Reinforcement Learning arises using the inherent properties of quantum mechanics, reducing the trainable parameters of a model significantly. However, gradient-based Multi-Agent Quantum Reinforcement Learning methods often have to struggle with barren plateaus, holding them back from matching the performance of classical approaches. We build upon an existing approach for gradient free Quantum Reinforcement Learning and propose three genetic variations with Variational Quantum Circuits for Multi-Agent Reinforcement Learning using evolutionary optimization. We evaluate our genetic variations in the Coin Game environment and also compare them to classical approaches. We showed that our Variational Quantum Circuit approaches perform significantly better compared to a neural network with a similar amount of trainable parameters. Compared to the larger neural network, our approaches archive similar results using $97.88\%$ less parameters.

Combining offline and online reinforcement learning (RL) is crucial for efficient and safe learning. However, previous approaches treat offline and online learning as separate procedures, resulting in redundant designs and limited performance. We ask: Can we achieve straightforward yet effective offline and online learning without introducing extra conservatism or regularization? In this study, we propose Uni-O4, which utilizes an on-policy objective for both offline and online learning. Owning to the alignment of objectives in two phases, the RL agent can transfer between offline and online learning seamlessly. This property enhances the flexibility of the learning paradigm, allowing for arbitrary combinations of pretraining, fine-tuning, offline, and online learning. In the offline phase, specifically, Uni-O4 leverages diverse ensemble policies to address the mismatch issues between the estimated behavior policy and the offline dataset. Through a simple offline policy evaluation (OPE) approach, Uni-O4 can achieve multi-step policy improvement safely. We demonstrate that by employing the method above, the fusion of these two paradigms can yield superior offline initialization as well as stable and rapid online fine-tuning capabilities. Through real-world robot tasks, we highlight the benefits of this paradigm for rapid deployment in challenging, previously unseen real-world environments. Additionally, through comprehensive evaluations using numerous simulated benchmarks, we substantiate that our method achieves state-of-the-art performance in both offline and offline-to-online fine-tuning learning. Imagine a scenario where a reinforcement learning robot needs to function and improve itself in the real world, the policy of the robot might go through the pipeline of training online in a simulator, then offline with real-world data, and lastly online in the real world. However, current reinforcement learning algorithms usually focus on specific stages of learning, which sophisticates the effort to train robots with a single unified framework. Online RL algorithms require a substantial amount of interaction and exploration to attain strong performance, which is prohibitive in many real-world applications. Offline RL, in which agents learn from a fixed dataset generated by other behavior policies, is a potential solution.

Train timetable rescheduling (TTR) aims to promptly restore the original operation of trains after unexpected disturbances or disruptions. Currently, this work is still done manually by train dispatchers, which is challenging to maintain performance under various problem instances. To mitigate this issue, this study proposes a reinforcement learning-based approach to TTR, which makes the following contributions compared to existing work. First, we design a simple directed graph to represent the TTR problem, enabling the automatic extraction of informative states through graph neural networks. Second, we reformulate the construction process of TTR's solution, not only decoupling the decision model from the problem size but also ensuring the generated scheme's feasibility. Third, we design a learning curriculum for our model to handle the scenarios with different levels of delay. Finally, a simple local search method is proposed to assist the learned decision model, which can significantly improve solution quality with little additional computation cost, further enhancing the practical value of our method. Extensive experimental results demonstrate the effectiveness of our method. The learned decision model can achieve better performance for various problems with varying degrees of train delay and different scales when compared to handcrafted rules and state-of-the-art solvers.