Cultural accumulation drives the open-ended and diverse progress in capabilities spanning human history. It builds an expanding body of knowledge and skills by combining individual exploration with inter-generational information transmission. Despite its widespread success among humans, the capacity for artificial learning agents to accumulate culture remains under-explored. In particular, approaches to reinforcement learning typically strive for improvements over only a single lifetime. Generational algorithms that do exist fail to capture the open-ended, emergent nature of cultural accumulation, which allows individuals to trade-off innovation and imitation. Building on the previously demonstrated ability for reinforcement learning agents to perform social learning, we find that training setups which balance this with independent learning give rise to cultural accumulation. These accumulating agents outperform those trained for a single lifetime with the same cumulative experience. We explore this accumulation by constructing two models under two distinct notions of a generation: episodic generations, in which accumulation occurs via in-context learning and train-time generations, in which accumulation occurs via in-weights learning. In-context and in-weights cultural accumulation can be interpreted as analogous to knowledge and skill accumulation, respectively. To the best of our knowledge, this work is the first to present general models that achieve emergent cultural accumulation in reinforcement learning, opening up new avenues towards more open-ended learning systems, as well as presenting new opportunities for modelling human culture.
In this work, we study the problem of finding Pareto optimal policies in multi-agent reinforcement learning problems with cooperative reward structures. We show that any algorithm where each agent only optimizes their reward is subject to suboptimal convergence. Therefore, to achieve Pareto optimality, agents have to act altruistically by considering the rewards of others. This observation bridges the multi-objective optimization framework and multi-agent reinforcement learning together. We first propose a framework for applying the Multiple Gradient Descent algorithm (MGDA) for learning in multi-agent settings. We further show that standard MGDA is subjected to weak Pareto convergence, a problem that is often overlooked in other learning settings but is prevalent in multi-agent reinforcement learning. To mitigate this issue, we propose MGDA++, an improvement of the existing algorithm to handle the weakly optimal convergence of MGDA properly. Theoretically, we prove that MGDA++ converges to strong Pareto optimal solutions in convex, smooth bi-objective problems. We further demonstrate the superiority of our MGDA++ in cooperative settings in the Gridworld benchmark. The results highlight that our proposed method can converge efficiently and outperform the other methods in terms of the optimality of the convergent policies. The source code is available at \url{https://github.com/giangbang/Strong-Pareto-MARL}.
Coordination among connected and autonomous vehicles (CAVs) is advancing due to developments in control and communication technologies. However, much of the current work is based on oversimplified and unrealistic task-specific assumptions, which may introduce vulnerabilities. This is critical because CAVs not only interact with their environment but are also integral parts of it. Insufficient exploration can result in policies that carry latent risks, highlighting the need for methods that explore the environment both extensively and efficiently. This work introduces OPTIMA, a novel distributed reinforcement learning framework for cooperative autonomous vehicle tasks. OPTIMA alternates between thorough data sampling from environmental interactions and multi-agent reinforcement learning algorithms to optimize CAV cooperation, emphasizing both safety and efficiency. Our goal is to improve the generality and performance of CAVs in highly complex and crowded scenarios. Furthermore, the industrial-scale distributed training system easily adapts to different algorithms, reward functions, and strategies.
Reinforcement learning algorithms for mean-field games offer a scalable framework for optimizing policies in large populations of interacting agents. Existing methods often depend on online interactions or access to system dynamics, limiting their practicality in real-world scenarios where such interactions are infeasible or difficult to model. In this paper, we present Offline Munchausen Mirror Descent (Off-MMD), a novel mean-field RL algorithm that approximates equilibrium policies in mean-field games using purely offline data. By leveraging iterative mirror descent and importance sampling techniques, Off-MMD estimates the mean-field distribution from static datasets without relying on simulation or environment dynamics. Additionally, we incorporate techniques from offline reinforcement learning to address common issues like Q-value overestimation, ensuring robust policy learning even with limited data coverage. Our algorithm scales to complex environments and demonstrates strong performance on benchmark tasks like crowd exploration or navigation, highlighting its applicability to real-world multi-agent systems where online experimentation is infeasible. We empirically demonstrate the robustness of Off-MMD to low-quality datasets and conduct experiments to investigate its sensitivity to hyperparameter choices.
A significant challenge for autonomous cyber defence is ensuring a defensive agent's ability to generalise across diverse network topologies and configurations. This capability is necessary for agents to remain effective when deployed in dynamically changing environments, such as an enterprise network where devices may frequently join and leave. Standard approaches to deep reinforcement learning, where policies are parameterised using a fixed-input multi-layer perceptron (MLP) expect fixed-size observation and action spaces. In autonomous cyber defence, this makes it hard to develop agents that generalise to environments with network topologies different from those trained on, as the number of nodes affects the natural size of the observation and action spaces. To overcome this limitation, we reframe the problem of autonomous network defence using entity-based reinforcement learning, where the observation and action space of an agent are decomposed into a collection of discrete entities. This framework enables the use of policy parameterisations specialised in compositional generalisation. We train a Transformer-based policy on the Yawning Titan cyber-security simulation environment and test its generalisation capabilities across various network topologies. We demonstrate that this approach significantly outperforms an MLP-based policy when training across fixed-size networks of varying topologies, and matches performance when training on a single network. We also demonstrate the potential for zero-shot generalisation to networks of a different size to those seen in training. These findings highlight the potential for entity-based reinforcement learning to advance the field of autonomous cyber defence by providing more generalisable policies capable of handling variations in real-world network environments.
The universe involves many independent co-learning agents as an ever-evolving part of our observed environment. Yet, in practice, Multi-Agent Reinforcement Learning (MARL) applications are typically constrained to small, homogeneous populations and remain computationally intensive. We propose a methodology that enables simulating populations of Reinforcement Learning agents at evolutionary scale. More specifically, we derive a fast, parallelizable implementation of Policy Gradient (PG) and Opponent-Learning Awareness (LOLA), tailored for evolutionary simulations where agents undergo random pairwise interactions in stateless normal-form games. We demonstrate our approach by simulating the evolution of very large populations made of heterogeneous co-learning agents, under both naive and advanced learning strategies. In our experiments, 200,000 PG or LOLA agents evolve in the classic games of Hawk-Dove, Stag-Hunt, and Rock-Paper-Scissors. Each game provides distinct insights into how populations evolve under both naive and advanced MARL rules, including compelling ways in which Opponent-Learning Awareness affects social evolution.
The correct specification of reward models is a well-known challenge in reinforcement learning. Hand-crafted reward functions often lead to inefficient or suboptimal policies and may not be aligned with user values. Reinforcement learning from human feedback is a successful technique that can mitigate such issues, however, the collection of human feedback can be laborious. Recent works have solicited feedback from pre-trained large language models rather than humans to reduce or eliminate human effort, however, these approaches yield poor performance in the presence of hallucination and other errors. This paper studies the advantages and limitations of reinforcement learning from large language model feedback and proposes a simple yet effective method for soliciting and applying feedback as a potential-based shaping function. We theoretically show that inconsistent rankings, which approximate ranking errors, lead to uninformative rewards with our approach. Our method empirically improves convergence speed and policy returns over commonly used baselines even with significant ranking errors, and eliminates the need for complex post-processing of reward functions.
Playing two-player games using reinforcement learning and self-play can be challenging due to the complexity of two-player environments and the possible instability in the training process. We propose that a reinforcement learning algorithm can train more efficiently and achieve improved performance in a two-player game if it leverages the knowledge from the single-player version of the same game. This study examines the proposed idea in ten different Atari 2600 environments using the Atari 2600 RAM as the input state. We discuss the advantages of using transfer learning from a single-player training process over training in a two-player setting from scratch, and demonstrate our results in a few measures such as training time and average total reward. We also discuss a method of calculating RAM complexity and its relationship to performance.
Proximal Policy Optimization (PPO) is commonly used in Reinforcement Learning from Human Feedback to align large language models (LLMs) with downstream tasks. This paper investigates the feasibility of using PPO for direct reinforcement learning (RL) from explicitly programmed reward signals, as opposed to indirect learning from human feedback via an intermediary reward model. We focus on tasks expressed through formal languages, such as mathematics and programming, where explicit reward functions can be programmed to automatically assess the quality of generated outputs. We apply this approach to a sentiment alignment task, a simple arithmetic task, and a more complex game synthesis task. The sentiment alignment task replicates prior research and serves to validate our experimental setup. Our results show that pure RL-based training for the two formal language tasks is challenging, with success being limited even for the simple arithmetic task. We propose a novel batch-entropy regularization term to aid exploration, although training is not yet entirely stable. Our findings suggest that direct RL training of LLMs may be more suitable for relatively minor changes, such as alignment, than for learning new tasks altogether, even if an informative reward signal can be expressed programmatically.
In recent years, reinforcement learning (RL) has gained popularity and has been applied to a wide range of tasks. One such popular domain where RL has been effective is resource management problems in systems. We look to extend work on RL for resource management problems by considering the novel domain of dynamic memory allocation management. We consider dynamic memory allocation to be a suitable domain for RL since current algorithms like first-fit, best-fit, and worst-fit can fail to adapt to changing conditions and can lead to fragmentation and suboptimal efficiency. In this paper, we present a framework in which an RL agent continuously learns from interactions with the system to improve memory management tactics. We evaluate our approach through various experiments using high-level and low-level action spaces and examine different memory allocation patterns. Our results show that RL can successfully train agents that can match and surpass traditional allocation strategies, particularly in environments characterized by adversarial request patterns. We also explore the potential of history-aware policies that leverage previous allocation requests to enhance the allocator's ability to handle complex request patterns. Overall, we find that RL offers a promising avenue for developing more adaptive and efficient memory allocation strategies, potentially overcoming limitations of hardcoded allocation algorithms.
Despite the popularity of multi-agent reinforcement learning (RL) in simulated and two-player applications, its success in messy real-world applications has been limited. A key challenge lies in its generalizability across problem variations, a common necessity for many real-world problems. Contextual reinforcement learning (CRL) formalizes learning policies that generalize across problem variations. However, the lack of standardized benchmarks for multi-agent CRL has hindered progress in the field. Such benchmarks are desired to be based on real-world applications to naturally capture the many open challenges of real-world problems that affect generalization. To bridge this gap, we propose IntersectionZoo, a comprehensive benchmark suite for multi-agent CRL through the real-world application of cooperative eco-driving in urban road networks. The task of cooperative eco-driving is to control a fleet of vehicles to reduce fleet-level vehicular emissions. By grounding IntersectionZoo in a real-world application, we naturally capture real-world problem characteristics, such as partial observability and multiple competing objectives. IntersectionZoo is built on data-informed simulations of 16,334 signalized intersections derived from 10 major US cities, modeled in an open-source industry-grade microscopic traffic simulator. By modeling factors affecting vehicular exhaust emissions (e.g., temperature, road conditions, travel demand), IntersectionZoo provides one million data-driven traffic scenarios. Using these traffic scenarios, we benchmark popular multi-agent RL and human-like driving algorithms and demonstrate that the popular multi-agent RL algorithms struggle to generalize in CRL settings.
Multi-agent systems are trained to maximize shared cost objectives, which typically reflect system-level efficiency. However, in the resource-constrained environments of mobility and transportation systems, efficiency may be achieved at the expense of fairness -- certain agents may incur significantly greater costs or lower rewards compared to others. Tasks could be distributed inequitably, leading to some agents receiving an unfair advantage while others incur disproportionately high costs. It is important to consider the tradeoffs between efficiency and fairness. We consider the problem of fair multi-agent navigation for a group of decentralized agents using multi-agent reinforcement learning (MARL). We consider the reciprocal of the coefficient of variation of the distances traveled by different agents as a measure of fairness and investigate whether agents can learn to be fair without significantly sacrificing efficiency (i.e., increasing the total distance traveled). We find that by training agents using min-max fair distance goal assignments along with a reward term that incentivizes fairness as they move towards their goals, the agents (1) learn a fair assignment of goals and (2) achieve almost perfect goal coverage in navigation scenarios using only local observations. For goal coverage scenarios, we find that, on average, our model yields a 14% improvement in efficiency and a 5% improvement in fairness over a baseline trained using random assignments. Furthermore, an average of 21% improvement in fairness can be achieved compared to a model trained on optimally efficient assignments; this increase in fairness comes at the expense of only a 7% decrease in efficiency. Finally, we extend our method to environments in which agents must complete coverage tasks in prescribed formations and show that it is possible to do so without tailoring the models to specific formation shapes.
Reinforcement learning (RL) faces challenges in trajectory planning for urban automated driving due to the poor convergence of RL and the difficulty in designing reward functions. The convergence problem is alleviated by combining RL with supervised learning. However, most existing approaches only reason one step ahead and lack the capability to plan for multiple future steps. Besides, although inverse reinforcement learning holds promise for solving the reward function design issue, existing methods for automated driving impose a linear structure assumption on reward functions, making them difficult to apply to urban automated driving. In light of these challenges, this paper proposes a novel RL-based trajectory planning method that integrates RL with imitation learning to enable multi-step planning. Furthermore, a transformer-based Bayesian reward function is developed, providing effective reward signals for RL in urban scenarios. Moreover, a hybrid-driven trajectory planning framework is proposed to enhance safety and interpretability. The proposed methods were validated on the large-scale real-world urban automated driving nuPlan dataset. The results demonstrated the significant superiority of the proposed methods over the baselines in terms of the closed-loop metrics. The code is available at https://github.com/Zigned/nuplan_zigned.
Chain-of-thought (CoT) reasoning in vision language models (VLMs) is crucial for improving interpretability and trustworthiness. However, current training recipes lack robust CoT reasoning data, relying on datasets dominated by short annotations with minimal rationales. In this work, we show that training VLM on short answers does not generalize well to reasoning tasks that require more detailed responses. To address this, we propose a two-fold approach. First, we distill rationales from GPT-4o model to enrich the training data and fine-tune VLMs, boosting their CoT performance. Second, we apply reinforcement learning to further calibrate reasoning quality. Specifically, we construct positive (correct) and negative (incorrect) pairs of model-generated reasoning chains, by comparing their predictions with annotated short answers. Using this pairwise data, we apply the Direct Preference Optimization algorithm to refine the model's reasoning abilities. Our experiments demonstrate significant improvements in CoT reasoning on benchmark datasets and better generalization to direct answer prediction as well. This work emphasizes the importance of incorporating detailed rationales in training and leveraging reinforcement learning to strengthen the reasoning capabilities of VLMs.
Deep Reinforcement learning (DRL) is used to enable autonomous navigation in unknown environments. Most research assume perfect sensor data, but real-world environments may contain natural and artificial sensor noise and denial. Here, we present a benchmark of both well-used and emerging DRL algorithms in a navigation task with configurable sensor denial effects. In particular, we are interested in comparing how different DRL methods (e.g. model-free PPO vs. model-based DreamerV3) are affected by sensor denial. We show that DreamerV3 outperforms other methods in the visual end-to-end navigation task with a dynamic goal - and other methods are not able to learn this. Furthermore, DreamerV3 generally outperforms other methods in sensor-denied environments. In order to improve robustness, we use adversarial training and demonstrate an improved performance in denied environments, although this generally comes with a performance cost on the vanilla environments. We anticipate this benchmark of different DRL methods and the usage of adversarial training to be a starting point for the development of more elaborate navigation strategies that are capable of dealing with uncertain and denied sensor readings.
This paper tackles the challenge of learning non-Markovian optimal execution strategies in dynamic financial markets. We introduce a novel actor-critic algorithm based on Deep Deterministic Policy Gradient (DDPG) to address this issue, with a focus on transient price impact modeled by a general decay kernel. Through numerical experiments with various decay kernels, we show that our algorithm successfully approximates the optimal execution strategy. Additionally, the proposed algorithm demonstrates adaptability to evolving market conditions, where parameters fluctuate over time. Our findings also show that modern reinforcement learning algorithms can provide a solution that reduces the need for frequent and inefficient human intervention in optimal execution tasks.
We demonstrate that recent advances in reinforcement learning (RL) combined with simple architectural changes significantly improves generalization on the ProcGen benchmark. These changes are frame stacking, replacing 2D convolutional layers with 3D convolutional layers, and scaling up the number of convolutional kernels per layer. Experimental results using a single set of hyperparameters across all environments show a 37.9\% reduction in the optimality gap compared to the baseline (from 0.58 to 0.36). This performance matches or exceeds current state-of-the-art methods. The proposed changes are largely orthogonal and therefore complementary to the existing approaches for improving generalization in RL, and our results suggest that further exploration in this direction could yield substantial improvements in addressing generalization challenges in deep reinforcement learning.
Traditionally, model-based reinforcement learning (MBRL) methods exploit neural networks as flexible function approximators to represent $\textit{a priori}$ unknown environment dynamics. However, training data are typically scarce in practice, and these black-box models often fail to generalize. Modeling architectures that leverage known physics can substantially reduce the complexity of system-identification, but break down in the face of complex phenomena such as contact. We introduce a novel framework for learning semi-structured dynamics models for contact-rich systems which seamlessly integrates structured first principles modeling techniques with black-box auto-regressive models. Specifically, we develop an ensemble of probabilistic models to estimate external forces, conditioned on historical observations and actions, and integrate these predictions using known Lagrangian dynamics. With this semi-structured approach, we can make accurate long-horizon predictions with substantially less data than prior methods. We leverage this capability and propose Semi-Structured Reinforcement Learning ($\texttt{SSRL}$) a simple model-based learning framework which pushes the sample complexity boundary for real-world learning. We validate our approach on a real-world Unitree Go1 quadruped robot, learning dynamic gaits -- from scratch -- on both hard and soft surfaces with just a few minutes of real-world data. Video and code are available at: https://sites.google.com/utexas.edu/ssrl
The emerging zero-shot capabilities of Large Language Models (LLMs) have led to their applications in areas extending well beyond natural language processing tasks. In reinforcement learning, while LLMs have been extensively used in text-based environments, their integration with continuous state spaces remains understudied. In this paper, we investigate how pre-trained LLMs can be leveraged to predict in context the dynamics of continuous Markov decision processes. We identify handling multivariate data and incorporating the control signal as key challenges that limit the potential of LLMs' deployment in this setup and propose Disentangled In-Context Learning (DICL) to address them. We present proof-of-concept applications in two reinforcement learning settings: model-based policy evaluation and data-augmented off-policy reinforcement learning, supported by theoretical analysis of the proposed methods. Our experiments further demonstrate that our approach produces well-calibrated uncertainty estimates. We release the code at https://github.com/abenechehab/dicl.
Constrained synthesizability is an unaddressed challenge in generative molecular design. In particular, designing molecules satisfying multi-parameter optimization objectives, while simultaneously being synthesizable and enforcing the presence of specific commercial building blocks in the synthesis. This is practically important for molecule re-purposing, sustainability, and efficiency. In this work, we propose a novel reward function called TANimoto Group Overlap (TANGO), which uses chemistry principles to transform a sparse reward function into a dense and learnable reward function -- crucial for reinforcement learning. TANGO can augment general-purpose molecular generative models to directly optimize for constrained synthesizability while simultaneously optimizing for other properties relevant to drug discovery using reinforcement learning. Our framework is general and addresses starting-material, intermediate, and divergent synthesis constraints. Contrary to most existing works in the field, we show that incentivizing a general-purpose (without any inductive biases) model is a productive approach to navigating challenging optimization scenarios. We demonstrate this by showing that the trained models explicitly learn a desirable distribution. Our framework is the first generative approach to tackle constrained synthesizability.