Oberman gave a stochastic control formulation of the problem of estimating the convex envelope of a non-convex function. Based on this, we develop a reinforcement learning scheme to approximate the convex envelope, using a variant of Q-learning for controlled optimal stopping. It shows very promising results on a standard library of test problems.

Prioritized Experience Replay (PER) enables the model to learn more about relatively important samples by artificially changing their accessed frequencies. However, this non-uniform sampling method shifts the state-action distribution that is originally used to estimate Q-value functions, which brings about the estimation deviation. In this article, an novel off policy reinforcement learning training framework called Directly Attention Loss Adjusted Prioritized Experience Replay (DALAP) is proposed, which can directly quantify the changed extent of the shifted distribution through Parallel Self-Attention network, so as to accurately compensate the error. In addition, a Priority-Encouragement mechanism is designed simultaneously to optimize the sample screening criterion, and further improve the training efficiency. In order to verify the effectiveness and generality of DALAP, we integrate it with the value-function based, the policy-gradient based and multi-agent reinforcement learning algorithm, respectively. The multiple groups of comparative experiments show that DALAP has the significant advantages of both improving the convergence rate and reducing the training variance.

Deep Learning (DL) is penetrating into a diverse range of mass mobility, smart living, and industrial applications, rapidly transforming the way we live and work. DL is at the heart of many AI implementations. A key set of challenges is to produce AI modules that are: (1) "circular" - can solve new tasks without forgetting how to solve previous ones, (2) "secure" - have immunity to adversarial data attacks, and (3) "tiny" - implementable in low power low cost embedded hardware. Clearly it is difficult to achieve all three aspects on a single horizontal layer of platforms, as the techniques require transformed deep representations that incur different computation and communication requirements. Here we set out the vision to achieve transformed DL representations across a 5G and Beyond networked architecture. We first detail the cross-sectoral motivations for each challenge area, before demonstrating recent advances in DL research that can achieve circular, secure, and tiny AI (CST-AI). Recognising the conflicting demand of each transformed deep representation, we federate their deep learning transformations and functionalities across the network to achieve connected run-time capabilities.

This paper presents a novel approach to generalizing robot manipulation skills by combining a sampling-based task-and-motion planner with an offline reinforcement learning algorithm. Starting with a small library of scripted primitive skills (e.g. Push) and object-centric symbolic predicates (e.g. On(block, plate)), the planner autonomously generates a demonstration dataset of manipulation skills in the context of a long-horizon task. An offline reinforcement learning algorithm then extracts a policy from the dataset without further interactions with the environment and replaces the scripted skill in the existing library. Refining the skill library improves the robustness of the planner, which in turn facilitates data collection for more complex manipulation skills. We validate our approach in simulation, on a block-pushing task. We show that the proposed method requires less training data than conventional reinforcement learning methods. Furthermore, interaction with the environment is collision-free because of the use of planner demonstrations, making the approach more amenable to persistent robot learning in the real world.

Despite the considerable potential of reinforcement learning (RL), robotics control tasks predominantly rely on imitation learning (IL) owing to its better sample efficiency. However, given the high cost of collecting extensive demonstrations, RL is still appealing if it can utilize limited imitation data for efficient autonomous self-improvement. Existing RL methods that utilize demonstrations either initialize the replay buffer with demonstrations and oversample them during RL training, which does not benefit from the generalization potential of modern IL methods, or pretrain the RL policy with IL on the demonstrations, which requires additional mechanisms to prevent catastrophic forgetting during RL fine-tuning. We propose imitation bootstrapped reinforcement learning (IBRL), a novel framework that first trains an IL policy on a limited number of demonstrations and then uses it to propose alternative actions for both online exploration and target value bootstrapping. IBRL achieves SoTA performance and sample efficiency on 7 challenging sparse reward continuous control tasks in simulation while learning directly from pixels. As a highlight of our method, IBRL achieves $6.4\times$ higher success rate than RLPD, a strong method that combines the idea of oversampling demonstrations with modern RL improvements, under the budget of 10 demos and 100K interactions in the challenging PickPlaceCan task in the Robomimic benchmark.

Experimentation on real robots is demanding in terms of time and costs. For this reason, a large part of the reinforcement learning (RL) community uses simulators to develop and benchmark algorithms. However, insights gained in simulation do not necessarily translate to real robots, in particular for tasks involving complex interactions with the environment. The Real Robot Challenge 2022 therefore served as a bridge between the RL and robotics communities by allowing participants to experiment remotely with a real robot - as easily as in simulation. In the last years, offline reinforcement learning has matured into a promising paradigm for learning from pre-collected datasets, alleviating the reliance on expensive online interactions. We therefore asked the participants to learn two dexterous manipulation tasks involving pushing, grasping, and in-hand orientation from provided real-robot datasets. An extensive software documentation and an initial stage based on a simulation of the real set-up made the competition particularly accessible. By giving each team plenty of access budget to evaluate their offline-learned policies on a cluster of seven identical real TriFinger platforms, we organized an exciting competition for machine learners and roboticists alike. In this work we state the rules of the competition, present the methods used by the winning teams and compare their results with a benchmark of state-of-the-art offline RL algorithms on the challenge datasets.

This paper introduces an effective and practical step toward approximate Bayesian inference in on-policy actor-critic deep reinforcement learning. This step manifests as three simple modifications to the Asynchronous Advantage Actor-Critic (A3C) algorithm: (1) applying a ReLU function to advantage estimates, (2) spectral normalization of actor-critic weights, and (3) incorporating dropout as a Bayesian approximation. We prove under standard assumptions that restricting policy updates to positive advantages optimizes for value by maximizing a lower bound on the value function plus an additive term. We show that the additive term is bounded proportional to the Lipschitz constant of the value function, which offers theoretical grounding for spectral normalization of critic weights. Finally, our application of dropout corresponds to approximate Bayesian inference over both the actor and critic parameters, which enables prudent state-aware exploration around the modes of the actor via Thompson sampling. Extensive empirical evaluations on diverse benchmarks reveal the superior performance of our approach compared to existing on- and off-policy algorithms. We demonstrate significant improvements for median and interquartile mean metrics over PPO, SAC, and TD3 on the MuJoCo continuous control benchmark. Moreover, we see improvement over PPO in the challenging ProcGen generalization benchmark.

Reward functions are difficult to design and often hard to align with human intent. Preference-based Reinforcement Learning (RL) algorithms address these problems by learning reward functions from human feedback. However, the majority of preference-based RL methods na\"ively combine supervised reward models with off-the-shelf RL algorithms. Contemporary approaches have sought to improve performance and query complexity by using larger and more complex reward architectures such as transformers. Instead of using highly complex architectures, we develop a new and parameter-efficient algorithm, Inverse Preference Learning (IPL), specifically designed for learning from offline preference data. Our key insight is that for a fixed policy, the $Q$-function encodes all information about the reward function, effectively making them interchangeable. Using this insight, we completely eliminate the need for a learned reward function. Our resulting algorithm is simpler and more parameter-efficient. Across a suite of continuous control and robotics benchmarks, IPL attains competitive performance compared to more complex approaches that leverage transformer-based and non-Markovian reward functions while having fewer algorithmic hyperparameters and learned network parameters. Our code is publicly released.

Recent breakthroughs in the development of agents to solve challenging sequential decision making problems such as Go [50], StarCraft [58], or DOTA [3], have relied on both simulated environments and large-scale datasets. However, progress on this research has been hindered by the scarcity of open-sourced datasets and the prohibitive computational cost to work with them. Here we present the NetHack Learning Dataset (NLD), a large and highly-scalable dataset of trajectories from the popular game of NetHack, which is both extremely challenging for current methods and very fast to run [23]. NLD consists of three parts: 10 billion state transitions from 1.5 million human trajectories collected on the NAO public NetHack server from 2009 to 2020; 3 billion state-action-score transitions from 100,000 trajectories collected from the symbolic bot winner of the NetHack Challenge 2021; and, accompanying code for users to record, load and stream any collection of such trajectories in a highly compressed form. We evaluate a wide range of existing algorithms including online and offline RL, as well as learning from demonstrations, showing that significant research advances are needed to fully leverage large-scale datasets for challenging sequential decision making tasks.

Cyclists prefer to use infrastructure that separates them from motorized traffic. Using a traffic light to segregate car and bike flows, with the addition of bike-specific green phases, is a lightweight and cheap solution that can be deployed dynamically to assess the opportunity of a heavier infrastructure such as a separate bike lane. To compensate for the increased waiting time induced by these new phases, we introduce in this paper a deep reinforcement learning solution that adapts the green phase cycle of a traffic light to the traffic. Vehicle counter data are used to compare the DRL approach with the actuated traffic light control algorithm over whole days. Results show that DRL achieves better minimization of vehicle waiting time at almost all hours. Our DRL approach is also robust to moderate changes in bike traffic. The code of this paper is available at https://github.com/LucasMagnana/A-DRL-solution-to-help-reduce-the-cost-in-waiting-time-of-securing-a-traffic-light-for-cyclists.