Obtaining reliable, adaptive confidence sets for prediction functions (hypotheses) is a central challenge in sequential decision-making tasks, such as bandits and model-based reinforcement learning. These confidence sets typically rely on prior assumptions on the hypothesis space, e.g., the known kernel of a Reproducing Kernel Hilbert Space (RKHS). Hand-designing such kernels is error prone, and misspecification may lead to poor or unsafe performance. In this work, we propose to meta-learn a kernel from offline data (Meta-KeL). For the case where the unknown kernel is a combination of known base kernels, we develop an estimator based on structured sparsity. Under mild conditions, we guarantee that our estimated RKHS yields valid confidence sets that, with increasing amounts of offline data, become as tight as those given the true unknown kernel. We demonstrate our approach on the kernelized bandit problem (a.k.a.~Bayesian optimization), where we establish regret bounds competitive with those given the true kernel. We also empirically evaluate the effectiveness of our approach on a Bayesian optimization task.

In this article, we will be going through what is Dueling Double Deep Q Learning and how to implement it in Tenroflow. Dueling Double Deep Q learning is the combination of Dueling Deep Q Learning and Double Deep Q Learning. Let's try to understand what is Dueling Deep Q learning and Double Deep Q Learning. One of the drawbacks of the DQN algorithm is that it overestimates the true rewards; the Q-values think the agent is going to obtain a higher return than what it will obtain in reality. This overestimation is due to the presence of Max of Q value for the next state in the Q learning update equation.

The target of reducing travel time only is insufficient to support the development of future smart transportation systems. To align with the United Nations Sustainable Development Goals (UN-SDG), a further reduction of fuel and emissions, improvements of traffic safety, and the ease of infrastructure deployment and maintenance should also be considered. Different from existing work focusing on the optimization of the control in either traffic light signal (to improve the intersection throughput), or vehicle speed (to stabilize the traffic), this paper presents a multi-agent deep reinforcement learning (DRL) system called CoTV, which Cooperatively controls both Traffic light signals and connected autonomous Vehicles (CAV). Therefore, our CoTV can well balance the achievement of the reduction of travel time, fuel, and emission. In the meantime, CoTV can also be easy to deploy by cooperating with only one CAV that is the nearest to the traffic light controller on each incoming road. This enables more efficient coordination between traffic light controllers and CAV, thus leading to the convergence of training CoTV under the large-scale multi-agent scenario that is traditionally difficult to converge. We give the detailed system design of CoTV, and demonstrate its effectiveness in a simulation study using SUMO under various grid maps and realistic urban scenarios with mixed-autonomy traffic.

We introduce Contrastive Intrinsic Control (CIC), an algorithm for unsupervised skill discovery that maximizes the mutual information between skills and state transitions. In contrast to most prior approaches, CIC uses a decomposition of the mutual information that explicitly incentivizes diverse behaviors by maximizing state entropy. We derive a novel lower bound estimate for the mutual information which combines a particle estimator for state entropy to generate diverse behaviors and contrastive learning to distill these behaviors into distinct skills. We evaluate our algorithm on the Unsupervised Reinforcement Learning Benchmark, which consists of a long reward-free pre-training phase followed by a short adaptation phase to downstream tasks with extrinsic rewards. We find that CIC substantially improves over prior unsupervised skill discovery methods and outperforms the next leading overall exploration algorithm in terms of downstream task performance.

AlphaGo, AlphaGo Zero, and all of their derivatives can play with superhuman strength because they are able to predict the win-lose outcome with great accuracy. However, Go as a game is decided by a final score difference, and in final positions AlphaGo plays suboptimal moves: this is not surprising, since AlphaGo is completely unaware of the final score difference, all winning final positions being equivalent from the winrate perspective. This can be an issue, for instance when trying to learn the "best" move or to play with an initial handicap. Moreover, there is the theoretical quest of the "perfect game", that is, the minimax solution. Thus, a natural question arises: is it possible to train a successful Reinforcement Learning agent to predict score differences instead of winrates? No empirical or theoretical evidence can be found in the literature to support the folklore statement that "this does not work". In this paper we present Leela Zero Score, a software designed to support or disprove the "does not work" statement. Leela Zero Score is designed on the open-source solution known as Leela Zero, and is trained on a 9x9 board to predict score differences instead of winrates. We find that the training produces a rational player, and we analyze its style against a strong amateur human player, to find that it is prone to some mistakes when the outcome is close. We compare its strength against SAI, an AlphaGo Zero-like software working on the 9x9 board, and find that the training of Leela Zero Score has reached a premature convergence to a player weaker than SAI.

Policy gradient (PG) estimation becomes a challenge when we are not allowed to sample with the target policy but only have access to a dataset generated by some unknown behavior policy. Conventional methods for off-policy PG estimation often suffer from either significant bias or exponentially large variance. In this paper, we propose the double Fitted PG estimation (FPG) algorithm. FPG can work with an arbitrary policy parameterization, assuming access to a Bellman-complete value function class. In the case of linear value function approximation, we provide a tight finite-sample upper bound on policy gradient estimation error, that is governed by the amount of distribution mismatch measured in feature space. We also establish the asymptotic normality of FPG estimation error with a precise covariance characterization, which is further shown to be statistically optimal with a matching Cramer-Rao lower bound. Empirically, we evaluate the performance of FPG on both policy gradient estimation and policy optimization, using either softmax tabular or ReLU policy networks. Under various metrics, our results show that FPG significantly outperforms existing off-policy PG estimation methods based on importance sampling and variance reduction techniques.

Although reinforcement learning has seen remarkable progress over the last years, solving robust dexterous object-manipulation tasks in multi-object settings remains a challenge. In this paper, we focus on models that can learn manipulation tasks in fixed multi-object settings and extrapolate this skill zero-shot without any drop in performance when the number of objects changes. We consider the generic task of bringing a specific cube out of a set to a goal position. We find that previous approaches, which primarily leverage attention and graph neural network-based architectures, do not generalize their skills when the number of input objects changes while scaling as $K^2$. We propose an alternative plug-and-play module based on relational inductive biases to overcome these limitations. Besides exceeding performances in their training environment, we show that our approach, which scales linearly in $K$, allows agents to extrapolate and generalize zero-shot to any new object number.

Generative Flow Networks (GFlowNets) are a method for learning a stochastic policy for generating compositional objects, such as graphs or strings, from a given unnormalized density by sequences of actions, where many possible action sequences may lead to the same object. Prior temporal difference-like learning objectives for training GFlowNets, such as flow matching and detailed balance, are prone to inefficient credit propagation across action sequences, particularly in the case of long sequences. We propose a new learning objective for GFlowNets, trajectory balance, as a more efficient alternative to previously used objectives. We prove that any global minimizer of the trajectory balance objective can define a policy that samples exactly from the target distribution. In experiments on four distinct domains, we empirically demonstrate the benefits of the trajectory balance objective for GFlowNet convergence, diversity of generated samples, and robustness to long action sequences and large action spaces.

We designed a Cloud-Based RL algorithm that continuously learns how to optimize power consumption by remotely reading the environmental data and consequently defining the HVAC set-points. In essence, MPC can fit complex thermodynamics and achieve excellent results in terms of energy savings on a single building. The main drawback of RBC is that they are difficult to be optimally tuned because they are not adaptable enough for the intrinsic complexity of the coupled building and plant thermodynamics. Therefore, it is desirable to introduce RL controls for large-scale applications on HVAC systems where the operating cost is high, like those in charge of the thermo-regulation of a significant volume. Supermarkets are, by definition, widespread buildings with variable thermal loads and complex occupational patterns that introduce a non-negligible stochastic component from the HVAC control point of view.

This paper presents a framework for learning visual representations from unlabeled video demonstrations captured from multiple viewpoints. We show that these representations are applicable for imitating several robotic tasks, including pick and place. We optimize a recently proposed self-supervised learning algorithm by applying contrastive learning to enhance task-relevant information while suppressing irrelevant information in the feature embeddings. We validate the proposed method on the publicly available Multi-View Pouring and a custom Pick and Place data sets and compare it with the TCN triplet baseline. We evaluate the learned representations using three metrics: viewpoint alignment, stage classification and reinforcement learning, and in all cases the results improve when compared to state-of-the-art approaches, with the added benefit of reduced number of training iterations.