Various data augmentation techniques have been recently proposed in image-based deep reinforcement learning (DRL). Although they empirically demonstrate the effectiveness of data augmentation for improving sample efficiency or generalization, which technique should be preferred is not always clear. To tackle this question, we analyze existing methods to better understand them and to uncover how they are connected. Notably, by expressing the variance of the Q-targets and that of the empirical actor/critic losses of these methods, we can analyze the effects of their different components and compare them. We furthermore formulate an explanation about how these methods may be affected by choosing different data augmentation transformations in calculating the target Q-values. This analysis suggests recommendations on how to exploit data augmentation in a more principled way. In addition, we include a regularization term called tangent prop, previously proposed in computer vision, but whose adaptation to DRL is novel to the best of our knowledge. We evaluate our proposition and validate our analysis in several domains. Compared to different relevant baselines, we demonstrate that it achieves state-of-the-art performance in most environments and shows higher sample efficiency and better generalization ability in some complex environments.
Building on a previous foundation work (Lussange et al. 2020), this study introduces a multi-agent reinforcement learning (MARL) model simulating crypto markets, which is calibrated to the Binance's daily closing prices of $153$ cryptocurrencies that were continuously traded between 2018 and 2022. Unlike previous agent-based models (ABM) or multi-agent systems (MAS) which relied on zero-intelligence agents or single autonomous agent methodologies, our approach relies on endowing agents with reinforcement learning (RL) techniques in order to model crypto markets. This integration is designed to emulate, with a bottom-up approach to complexity inference, both individual and collective agents, ensuring robustness in the recent volatile conditions of such markets and during the COVID-19 era. A key feature of our model also lies in the fact that its autonomous agents perform asset price valuation based on two sources of information: the market prices themselves, and the approximation of the crypto assets fundamental values beyond what those market prices are. Our MAS calibration against real market data allows for an accurate emulation of crypto markets microstructure and probing key market behaviors, in both the bearish and bullish regimes of that particular time period.
We study the robustness of deep reinforcement learning algorithms against distribution shifts within contextual multi-stage stochastic combinatorial optimization problems from the operations research domain. In this context, risk-sensitive algorithms promise to learn robust policies. While this field is of general interest to the reinforcement learning community, most studies up-to-date focus on theoretical results rather than real-world performance. With this work, we aim to bridge this gap by formally deriving a novel risk-sensitive deep reinforcement learning algorithm while providing numerical evidence for its efficacy. Specifically, we introduce discrete Soft Actor-Critic for the entropic risk measure by deriving a version of the Bellman equation for the respective Q-values. We establish a corresponding policy improvement result and infer a practical algorithm. We introduce an environment that represents typical contextual multi-stage stochastic combinatorial optimization problems and perform numerical experiments to empirically validate our algorithm's robustness against realistic distribution shifts, without compromising performance on the training distribution. We show that our algorithm is superior to risk-neutral Soft Actor-Critic as well as to two benchmark approaches for robust deep reinforcement learning. Thereby, we provide the first structured analysis on the robustness of reinforcement learning under distribution shifts in the realm of contextual multi-stage stochastic combinatorial optimization problems.
When Reinforcement Learning (RL) agents are deployed in practice, they might impact their environment and change its dynamics. We propose a new framework to model this phenomenon, where the current environment depends on the deployed policy as well as its previous dynamics. This is a generalization of Performative RL (PRL) [Mandal et al., 2023]. Unlike PRL, our framework allows to model scenarios where the environment gradually adjusts to a deployed policy. We adapt two algorithms from the performative prediction literature to our setting and propose a novel algorithm called Mixed Delayed Repeated Retraining (MDRR). We provide conditions under which these algorithms converge and compare them using three metrics: number of retrainings, approximation guarantee, and number of samples per deployment. MDRR is the first algorithm in this setting which combines samples from multiple deployments in its training. This makes MDRR particularly suitable for scenarios where the environment's response strongly depends on its previous dynamics, which are common in practice. We experimentally compare the algorithms using a simulation-based testbed and our results show that MDRR converges significantly faster than previous approaches.
The human-like automatic deductive reasoning has always been one of the most challenging open problems in the interdiscipline of mathematics and artificial intelligence. This paper is the third in a series of our works. We built a neural-symbolic system, called FGeoDRL, to automatically perform human-like geometric deductive reasoning. The neural part is an AI agent based on reinforcement learning, capable of autonomously learning problem-solving methods from the feedback of a formalized environment, without the need for human supervision. It leverages a pre-trained natural language model to establish a policy network for theorem selection and employ Monte Carlo Tree Search for heuristic exploration. The symbolic part is a reinforcement learning environment based on geometry formalization theory and FormalGeo, which models GPS as a Markov Decision Process. In this formal symbolic system, the known conditions and objectives of the problem form the state space, while the set of theorems forms the action space. Leveraging FGeoDRL, we have achieved readable and verifiable automated solutions to geometric problems. Experiments conducted on the formalgeo7k dataset have achieved a problem-solving success rate of 86.40%. The project is available at https://github.com/PersonNoName/FGeoDRL.
The inverse reinforcement learning approach to imitation learning is a double-edged sword. On the one hand, it can enable learning from a smaller number of expert demonstrations with more robustness to error compounding than behavioral cloning approaches. On the other hand, it requires that the learner repeatedly solve a computationally expensive reinforcement learning (RL) problem. Often, much of this computation is wasted searching over policies very dissimilar to the expert's. In this work, we propose using hybrid RL -- training on a mixture of online and expert data -- to curtail unnecessary exploration. Intuitively, the expert data focuses the learner on good states during training, which reduces the amount of exploration required to compute a strong policy. Notably, such an approach doesn't need the ability to reset the learner to arbitrary states in the environment, a requirement of prior work in efficient inverse RL. More formally, we derive a reduction from inverse RL to expert-competitive RL (rather than globally optimal RL) that allows us to dramatically reduce interaction during the inner policy search loop while maintaining the benefits of the IRL approach. This allows us to derive both model-free and model-based hybrid inverse RL algorithms with strong policy performance guarantees. Empirically, we find that our approaches are significantly more sample efficient than standard inverse RL and several other baselines on a suite of continuous control tasks.
Reinforcement learning (RL) has been widely adopted for controlling and optimizing complex engineering systems such as next-generation wireless networks. An important challenge in adopting RL is the need for direct access to the physical environment. This limitation is particularly severe in multi-agent systems, for which conventional multi-agent reinforcement learning (MARL) requires a large number of coordinated online interactions with the environment during training. When only offline data is available, a direct application of online MARL schemes would generally fail due to the epistemic uncertainty entailed by the lack of exploration during training. In this work, we propose an offline MARL scheme that integrates distributional RL and conservative Q-learning to address the environment's inherent aleatoric uncertainty and the epistemic uncertainty arising from the use of offline data. We explore both independent and joint learning strategies. The proposed MARL scheme, referred to as multi-agent conservative quantile regression, addresses general risk-sensitive design criteria and is applied to the trajectory planning problem in drone networks, showcasing its advantages.
The recent rapid progress in (self) supervised learning models is in large part predicted by empirical scaling laws: a model's performance scales proportionally to its size. Analogous scaling laws remain elusive for reinforcement learning domains, however, where increasing the parameter count of a model often hurts its final performance. In this paper, we demonstrate that incorporating Mixture-of-Expert (MoE) modules, and in particular Soft MoEs (Puigcerver et al., 2023), into value-based networks results in more parameter-scalable models, evidenced by substantial performance increases across a variety of training regimes and model sizes. This work thus provides strong empirical evidence towards developing scaling laws for reinforcement learning.
This paper contributes a new approach for distributional reinforcement learning which elucidates a clean separation of transition structure and reward in the learning process. Analogous to how the successor representation (SR) describes the expected consequences of behaving according to a given policy, our distributional successor measure (SM) describes the distributional consequences of this behaviour. We formulate the distributional SM as a distribution over distributions and provide theory connecting it with distributional and model-based reinforcement learning. Moreover, we propose an algorithm that learns the distributional SM from data by minimizing a two-level maximum mean discrepancy. Key to our method are a number of algorithmic techniques that are independently valuable for learning generative models of state. As an illustration of the usefulness of the distributional SM, we show that it enables zero-shot risk-sensitive policy evaluation in a way that was not previously possible.
Scripted agents have predominantly won the five previous iterations of the IEEE microRTS ($\mu$RTS) competitions hosted at CIG and CoG. Despite Deep Reinforcement Learning (DRL) algorithms making significant strides in real-time strategy (RTS) games, their adoption in this primarily academic competition has been limited due to the considerable training resources required and the complexity inherent in creating and debugging such agents. RAISocketAI is the first DRL agent to win the IEEE microRTS competition. In a benchmark without performance constraints, RAISocketAI regularly defeated the two prior competition winners. This first competition-winning DRL submission can be a benchmark for future microRTS competitions and a starting point for future DRL research. Iteratively fine-tuning the base policy and transfer learning to specific maps were critical to RAISocketAI's winning performance. These strategies can be used to economically train future DRL agents. Further work in Imitation Learning using Behavior Cloning and fine-tuning these models with DRL has proven promising as an efficient way to bootstrap models with demonstrated, competitive behaviors.
An important challenge in multi-objective reinforcement learning is obtaining a Pareto front of policies to attain optimal performance under different preferences. We introduce Iterated Pareto Referent Optimisation (IPRO), which decomposes finding the Pareto front into a sequence of constrained single-objective problems. This enables us to guarantee convergence while providing an upper bound on the distance to undiscovered Pareto optimal solutions at each step. We evaluate IPRO using utility-based metrics and its hypervolume and find that it matches or outperforms methods that require additional assumptions. By leveraging problem-specific single-objective solvers, our approach also holds promise for applications beyond multi-objective reinforcement learning, such as planning and pathfinding.
Reward functions are central in specifying the task we want a reinforcement learning agent to perform. Given a task and desired optimal behavior, we study the problem of designing informative reward functions so that the designed rewards speed up the agent's convergence. In particular, we consider expert-driven reward design settings where an expert or teacher seeks to provide informative and interpretable rewards to a learning agent. Existing works have considered several different reward design formulations; however, the key challenge is formulating a reward informativeness criterion that adapts w.r.t. the agent's current policy and can be optimized under specified structural constraints to obtain interpretable rewards. In this paper, we propose a novel reward informativeness criterion, a quantitative measure that captures how the agent's current policy will improve if it receives rewards from a specific reward function. We theoretically showcase the utility of the proposed informativeness criterion for adaptively designing rewards for an agent. Experimental results on two navigation tasks demonstrate the effectiveness of our adaptive reward informativeness criterion.
Although deep reinforcement learning methods can learn effective policies for challenging problems such as Atari games and robotics tasks, algorithms are complex, and training times are often long. This study investigates how Evolution Strategies perform compared to gradient-based deep reinforcement learning methods. We use Evolution Strategies to optimize the weights of a neural network via neuroevolution, performing direct policy search. We benchmark both deep policy networks and networks consisting of a single linear layer from observations to actions for three gradient-based methods, such as Proximal Policy Optimization. These methods are evaluated against three classical Evolution Strategies and Augmented Random Search, which all use linear policy networks. Our results reveal that Evolution Strategies can find effective linear policies for many reinforcement learning benchmark tasks, unlike deep reinforcement learning methods that can only find successful policies using much larger networks, suggesting that current benchmarks are easier to solve than previously assumed. Interestingly, Evolution Strategies also achieve results comparable to gradient-based deep reinforcement learning algorithms for higher-complexity tasks. Furthermore, we find that by directly accessing the memory state of the game, Evolution Strategies can find successful policies in Atari that outperform the policies found by Deep Q-Learning. Evolution Strategies also outperform Augmented Random Search in most benchmarks, demonstrating superior sample efficiency and robustness in training linear policy networks.
Deceptive path planning (DPP) is the problem of designing a path that hides its true goal from an outside observer. Existing methods for DPP rely on unrealistic assumptions, such as global state observability and perfect model knowledge, and are typically problem-specific, meaning that even minor changes to a previously solved problem can force expensive computation of an entirely new solution. Given these drawbacks, such methods do not generalize to unseen problem instances, lack scalability to realistic problem sizes, and preclude both on-the-fly tunability of deception levels and real-time adaptivity to changing environments. In this paper, we propose a reinforcement learning (RL)-based scheme for training policies to perform DPP over arbitrary weighted graphs that overcomes these issues. The core of our approach is the introduction of a local perception model for the agent, a new state space representation distilling the key components of the DPP problem, the use of graph neural network-based policies to facilitate generalization and scaling, and the introduction of new deception bonuses that translate the deception objectives of classical methods to the RL setting. Through extensive experimentation we show that, without additional fine-tuning, at test time the resulting policies successfully generalize, scale, enjoy tunable levels of deception, and adapt in real-time to changes in the environment.
This paper introduces a novel approach, Decision Theory-guided Deep Reinforcement Learning (DT-guided DRL), to address the inherent cold start problem in DRL. By integrating decision theory principles, DT-guided DRL enhances agents' initial performance and robustness in complex environments, enabling more efficient and reliable convergence during learning. Our investigation encompasses two primary problem contexts: the cart pole and maze navigation challenges. Experimental results demonstrate that the integration of decision theory not only facilitates effective initial guidance for DRL agents but also promotes a more structured and informed exploration strategy, particularly in environments characterized by large and intricate state spaces. The results of experiment demonstrate that DT-guided DRL can provide significantly higher rewards compared to regular DRL. Specifically, during the initial phase of training, the DT-guided DRL yields up to an 184% increase in accumulated reward. Moreover, even after reaching convergence, it maintains a superior performance, ending with up to 53% more reward than standard DRL in large maze problems. DT-guided DRL represents an advancement in mitigating a fundamental challenge of DRL by leveraging functions informed by human (designer) knowledge, setting a foundation for further research in this promising interdisciplinary domain.
Reinforcement learning (RL) is a classical tool to solve network control or policy optimization problems in unknown environments. The original Q-learning suffers from performance and complexity challenges across very large networks. Herein, a novel model-free ensemble reinforcement learning algorithm which adapts the classical Q-learning is proposed to handle these challenges for networks which admit Markov decision process (MDP) models. Multiple Q-learning algorithms are run on multiple, distinct, synthetically created and structurally related Markovian environments in parallel; the outputs are fused using an adaptive weighting mechanism based on the Jensen-Shannon divergence (JSD) to obtain an approximately optimal policy with low complexity. The theoretical justification of the algorithm, including the convergence of key statistics and Q-functions are provided. Numerical results across several network models show that the proposed algorithm can achieve up to 55% less average policy error with up to 50% less runtime complexity than the state-of-the-art Q-learning algorithms. Numerical results validate assumptions made in the theoretical analysis.
Recent advancements in meta-learning have enabled the automatic discovery of novel reinforcement learning algorithms parameterized by surrogate objective functions. To improve upon manually designed algorithms, the parameterization of this learned objective function must be expressive enough to represent novel principles of learning (instead of merely recovering already established ones) while still generalizing to a wide range of settings outside of its meta-training distribution. However, existing methods focus on discovering objective functions that, like many widely used objective functions in reinforcement learning, do not take into account the total number of steps allowed for training, or "training horizon". In contrast, humans use a plethora of different learning objectives across the course of acquiring a new ability. For instance, students may alter their studying techniques based on the proximity to exam deadlines and their self-assessed capabilities. This paper contends that ignoring the optimization time horizon significantly restricts the expressive potential of discovered learning algorithms. We propose a simple augmentation to two existing objective discovery approaches that allows the discovered algorithm to dynamically update its objective function throughout the agent's training procedure, resulting in expressive schedules and increased generalization across different training horizons. In the process, we find that commonly used meta-gradient approaches fail to discover such adaptive objective functions while evolution strategies discover highly dynamic learning rules. We demonstrate the effectiveness of our approach on a wide range of tasks and analyze the resulting learned algorithms, which we find effectively balance exploration and exploitation by modifying the structure of their learning rules throughout the agent's lifetime.
Inspired by the insights in cognitive science with respect to human memory and reasoning mechanism, a novel evolvable LLM-based (Large Language Model) agent framework is proposed as REMEMBERER. By equipping the LLM with a long-term experience memory, REMEMBERER is capable of exploiting the experiences from the past episodes even for different task goals, which excels an LLM-based agent with fixed exemplars or equipped with a transient working memory. We further introduce Reinforcement Learning with Experience Memory (RLEM) to update the memory. Thus, the whole system can learn from the experiences of both success and failure, and evolve its capability without fine-tuning the parameters of the LLM. In this way, the proposed REMEMBERER constitutes a semi-parametric RL agent. Extensive experiments are conducted on two RL task sets to evaluate the proposed framework. The average results with different initialization and training sets exceed the prior SOTA by 4% and 2% for the success rate on two task sets and demonstrate the superiority and robustness of REMEMBERER.
Graph-based environments pose unique challenges to multi-agent reinforcement learning. In decentralized approaches, agents operate within a given graph and make decisions based on partial or outdated observations. The size of the observed neighborhood limits the generalizability to different graphs and affects the reactivity of agents, the quality of the selected actions, and the communication overhead. This work focuses on generalizability and resolves the trade-off in observed neighborhood size with a continuous information flow in the whole graph. We propose a recurrent message-passing model that iterates with the environment's steps and allows nodes to create a global representation of the graph by exchanging messages with their neighbors. Agents receive the resulting learned graph observations based on their location in the graph. Our approach can be used in a decentralized manner at runtime and in combination with a reinforcement learning algorithm of choice. We evaluate our method across 1000 diverse graphs in the context of routing in communication networks and find that it enables agents to generalize and adapt to changes in the graph.
Reinforcement learning algorithms need exploration to learn. However, unsupervised exploration prevents the deployment of such algorithms on safety-critical tasks and limits real-world deployment. In this paper, we propose a new algorithm called Ensemble Model Predictive Safety Certification that combines model-based deep reinforcement learning with tube-based model predictive control to correct the actions taken by a learning agent, keeping safety constraint violations at a minimum through planning. Our approach aims to reduce the amount of prior knowledge about the actual system by requiring only offline data generated by a safe controller. Our results show that we can achieve significantly fewer constraint violations than comparable reinforcement learning methods.