The recent observation of neural power-law scaling relations has made a significant impact in the field of deep learning. A substantial amount of attention has been dedicated as a consequence to the description of scaling laws, although mostly for supervised learning and only to a reduced extent for reinforcement learning frameworks. In this paper we present an extensive study of performance scaling for a cornerstone reinforcement learning algorithm, AlphaZero. On the basis of a relationship between Elo rating, playing strength and power-law scaling, we train AlphaZero agents on the games Connect Four and Pentago and analyze their performance. We find that player strength scales as a power law in neural network parameter count when not bottlenecked by available compute, and as a power of compute when training optimally sized agents. We observe nearly identical scaling exponents for both games. Combining the two observed scaling laws we obtain a power law relating optimal size to compute similar to the ones observed for language models. We find that the predicted scaling of optimal neural network size fits our data for both games. We also show that large AlphaZero models are more sample efficient, performing better than smaller models with the same amount of training data. The range of fields investigated include natural language processing and computer vision (Rosenfeld et al., 2019). Most of these scaling laws regard the dependency of test loss on either dataset size, number of neural network parameters, or training compute. The robustness of the observed scaling laws across many orders of magnitude led to the creation of large models, with parameters numbering in the tens and hundreds of billions (Brown et al., 2020; Hoffmann et al., 2022; Alayrac et al., 2022). Until now, evidence for power-law scaling has come in most part from supervised learning methods. Considerably less effort has been dedicated to the scaling of reinforcement learning algorithms, such as performance scaling with model size (Reed et al., 2022; Lee et al., 2022).

In this work we address the problem of finding feasible policies for Constrained Markov Decision Processes under probability one constraints. We argue that stationary policies are not sufficient for solving this problem, and that a rich class of policies can be found by endowing the controller with a scalar quantity, so called budget, that tracks how close the agent is to violating the constraint. We show that the minimal budget required to act safely can be obtained as the smallest fixed point of a Bellman-like operator, for which we analyze its convergence properties. We also show how to learn this quantity when the true kernel of the Markov decision process is not known, while providing sample-complexity bounds. The utility of knowing this minimal budget relies in that it can aid in the search of optimal or near-optimal policies by shrinking down the region of the state space the agent must navigate. Simulations illustrate the different nature of probability one constraints against the typically used constraints in expectation.

To control the behavior of physics-based characters through language, Nvidia's PADL combines a language model with reinforcement learning. One thing that comes to mind when remembering the start of this wave of AI is certainly the strangely moving 3D figures from Deepmind and other research institutions. These three-legged spiders or humanoid 3D puppets had learned their movements through reinforcement learning. There are now numerous approaches to making digital animals or human-like figures learn movements on their own. The goal of these methods is to develop AI systems that can generate natural-looking movements for a variety of simulated figures and thus complement or replace manual animation and motion capture processes in the long term.

Continuous trajectory tracking control of quadrotors is complicated when considering noise from the environment. Due to the difficulty in modeling the environmental dynamics, tracking methodologies based on conventional control theory, such as model predictive control, have limitations on tracking accuracy and response time. We propose a Time-attenuating Twin Delayed DDPG, a model-free algorithm that is robust to noise, to better handle the trajectory tracking task. A deep reinforcement learning framework is constructed, where a time decay strategy is designed to avoid trapping into local optima. The experimental results show that the tracking error is significantly small, and the operation time is one-tenth of that of a traditional algorithm. The OpenAI Mujoco tool is used to verify the proposed algorithm, and the simulation results show that, the proposed method can significantly improve the training efficiency and effectively improve the accuracy and convergence stability.

With the continuous increment of maritime applications, the development of marine networks for data offloading becomes necessary. However, the limited maritime network resources are very difficult to satisfy real-time demands. Besides, how to effectively handle multiple compute-intensive tasks becomes another intractable issue. Hence, in this paper, we focus on the decision of maritime task offloading by the cooperation of unmanned aerial vehicles (UAVs) and vessels. Specifically, we first propose a cooperative offloading framework, including the demands from marine Internet of Things (MIoTs) devices and resource providers from UAVs and vessels. Due to the limited energy and computation ability of UAVs, it is necessary to help better apply the vessels to computation offloading. Then, we formulate the studied problem into a Markov decision process, aiming to minimize the total execution time and energy cost. Then, we leverage Lyapunov optimization to convert the long-term constraints of the total execution time and energy cost into their short-term constraints, further yielding a set of per-time-slot optimization problems. Furthermore, we propose a Q-learning based approach to solve the short-term problem efficiently. Finally, simulation results are conducted to verify the correctness and effectiveness of the proposed algorithm.

We study how to transfer representations pretrained on source tasks to target tasks in visual percept based RL. We analyze two popular approaches: freezing or finetuning the pretrained representations. Empirical studies on a set of popular tasks reveal several properties of pretrained representations. First, finetuning is required even when pretrained representations perfectly capture the information required to solve the target task. Second, finetuned representations improve learnability and are more robust to noise. Third, pretrained bottom layers are task-agnostic and readily transferable to new tasks, while top layers encode task-specific information and require adaptation. Building on these insights, we propose a self-supervised objective that clusters representations according to the policy they induce, as opposed to traditional representation similarity measures which are policy-agnostic (e.g. Euclidean norm, cosine similarity). Together with freezing the bottom layers, this objective results in significantly better representation than frozen, finetuned, and self-supervised alternatives on a wide range of benchmarks.

Recently, short video platforms have achieved rapid user growth by recommending interesting content to users. The objective of the recommendation is to optimize user retention, thereby driving the growth of DAU (Daily Active Users). Retention is a long-term feedback after multiple interactions of users and the system, and it is hard to decompose retention reward to each item or a list of items. Thus traditional point-wise and list-wise models are not able to optimize retention. In this paper, we choose reinforcement learning methods to optimize the retention as they are designed to maximize the long-term performance. We formulate the problem as an infinite-horizon request-based Markov Decision Process, and our objective is to minimize the accumulated time interval of multiple sessions, which is equal to improving the app open frequency and user retention. However, current reinforcement learning algorithms can not be directly applied in this setting due to uncertainty, bias, and long delay time incurred by the properties of user retention. We propose a novel method, dubbed RLUR, to address the aforementioned challenges. Both offline and live experiments show that RLUR can significantly improve user retention. RLUR has been fully launched in Kuaishou app for a long time, and achieves consistent performance improvement on user retention and DAU.

Inspired by recent progress in multi-agent Reinforcement Learning (RL), in this work we examine the collective intelligent behaviour of theoretical universal agents by introducing a weighted mixture operation. Given a weighted set of agents, their weighted mixture is a new agent whose expected total reward in any environment is the corresponding weighted average of the original agents' expected total rewards in that environment. Thus, if RL agent intelligence is quantified in terms of performance across environments, the weighted mixture's intelligence is the weighted average of the original agents' intelligences. This operation enables various interesting new theorems that shed light on the geometry of RL agent intelligence, namely: results about symmetries, convex agent-sets, and local extrema. We also show that any RL agent intelligence measure based on average performance across environments, subject to certain weak technical conditions, is identical (up to a constant factor) to performance within a single environment dependent on said intelligence measure.

The AAAI 2023 best paper awards were presented at the conference on Saturday 11 February. The awards comprised one outstanding paper, one outstanding student paper, and 12 distinguished papers. The AAAI outstanding paper award is given to a paper (or papers) that "exemplifies the highest standards in technical contribution and exposition". Abstract: The aim of Inverse Reinforcement Learning (IRL) is to infer a reward function R from a policy pi. To do this, we need a model of how pi relates to R. In the current literature, the most common models are optimality, Boltzmann rationality, and causal entropy maximisation.

Generative Flow Networks (GFlowNets) are a new family of probabilistic samplers where an agent learns a stochastic policy for generating complex combinatorial structure through a series of decision-making steps. Despite being inspired from reinforcement learning, the current GFlowNet framework is relatively limited in its applicability and cannot handle stochasticity in the reward function. In this work, we adopt a distributional paradigm for GFlowNets, turning each flow function into a distribution, thus providing more informative learning signals during training. By parameterizing each edge flow through their quantile functions, our proposed \textit{quantile matching} GFlowNet learning algorithm is able to learn a risk-sensitive policy, an essential component for handling scenarios with risk uncertainty. Moreover, we find that the distributional approach can achieve substantial improvement on existing benchmarks compared to prior methods due to our enhanced training algorithm, even in settings with deterministic rewards.