Complex reasoning problems contain states that vary in the computational cost required to determine a good action plan. Taking advantage of this property, we propose Adaptive Subgoal Search (AdaSubS), a search method that adaptively adjusts the planning horizon. To this end, AdaSubS generates diverse sets of subgoals at different distances. A verification mechanism is employed to filter out unreachable subgoals swiftly, allowing to focus on feasible further subgoals. In this way, AdaSubS benefits from the efficiency of planning with longer subgoals and the fine control with the shorter ones, and thus scales well to difficult planning problems. We show that AdaSubS significantly surpasses hierarchical planning algorithms on three complex reasoning tasks: Sokoban, the Rubik's Cube, and inequality proving benchmark INT.

Autonomous cyber and cyber-physical systems need to perform decision-making, learning, and control in unknown environments. Such decision-making can be sensitive to multiple factors, including modeling errors, changes in costs, and impacts of events in the tails of probability distributions. Although multi-agent reinforcement learning (MARL) provides a framework for learning behaviors through repeated interactions with the environment by minimizing an average cost, it will not be adequate to overcome the above challenges. In this paper, we develop a distributed MARL approach to solve decision-making problems in unknown environments by learning risk-aware actions. We use the conditional value-at-risk (CVaR) to characterize the cost function that is being minimized, and define a Bellman operator to characterize the value function associated to a given state-action pair. We prove that this operator satisfies a contraction property, and that it converges to the optimal value function. We then propose a distributed MARL algorithm called the CVaR QD-Learning algorithm, and establish that value functions of individual agents reaches consensus. We identify several challenges that arise in the implementation of the CVaR QD-Learning algorithm, and present solutions to overcome these. We evaluate the CVaR QD-Learning algorithm through simulations, and demonstrate the effect of a risk parameter on value functions at consensus.

The contribution of this paper is a generalized formulation of correctional learning using optimal transport, which is about how to optimally transport one mass distribution to another. Correctional learning is a framework developed to enhance the accuracy of parameter estimation processes by means of a teacher-student approach. In this framework, an expert agent, referred to as the teacher, modifies the data used by a learning agent, known as the student, to improve its estimation process. The objective of the teacher is to alter the data such that the student's estimation error is minimized, subject to a fixed intervention budget. Compared to existing formulations of correctional learning, our novel optimal transport approach provides several benefits. It allows for the estimation of more complex characteristics as well as the consideration of multiple intervention policies for the teacher. We evaluate our approach on two theoretical examples, and on a human-robot interaction application in which the teacher's role is to improve the robots performance in an inverse reinforcement learning setting.

Any quantum computing application, once encoded as a quantum circuit, must be compiled before being executable on a quantum computer. Similar to classical compilation, quantum compilation is a sequential process with many compilation steps and numerous possible optimization passes. Despite the similarities, the development of compilers for quantum computing is still in its infancy -- lacking mutual consolidation on the best sequence of passes, compatibility, adaptability, and flexibility. In this work, we take advantage of decades of classical compiler optimization and propose a reinforcement learning framework for developing optimized quantum circuit compilation flows. Through distinct constraints and a unifying interface, the framework supports the combination of techniques from different compilers and optimization tools in a single compilation flow. Experimental evaluations show that the proposed framework -- set up with a selection of compilation passes from IBM's Qiskit and Quantinuum's TKET -- significantly outperforms both individual compilers in 73% of cases regarding the expected fidelity. The framework is available on GitHub (https://github.com/cda-tum/MQTPredictor) as part of the Munich Quantum Toolkit (MQT).

We formulate an efficient approximation for multi-agent batch reinforcement learning, the approximated multi-agent fitted Q iteration (AMAFQI). We present a detailed derivation of our approach. We propose an iterative policy search and show that it yields a greedy policy with respect to multiple approximations of the centralized, learned Q-function. In each iteration and policy evaluation, AMAFQI requires a number of computations that scales linearly with the number of agents whereas the analogous number of computations increase exponentially for the fitted Q iteration (FQI), a commonly used approaches in batch reinforcement learning. This property of AMAFQI is fundamental for the design of a tractable multi-agent approach. We evaluate the performance of AMAFQI and compare it to FQI in numerical simulations. The simulations illustrate the significant computation time reduction when using AMAFQI instead of FQI in multi-agent problems and corroborate the similar performance of both approaches.

When using a tool, the grasps used for picking it up, reposing, and holding it in a suitable pose for the desired task could be distinct. Therefore, a key challenge for autonomous in-hand tool manipulation is finding a sequence of grasps that facilitates every step of the tool use process while continuously maintaining force closure and stability. Due to the complexity of modeling the contact dynamics, reinforcement learning (RL) techniques can provide a solution in this continuous space subject to highly parameterized physical models. However, these techniques impose a trade-off in adaptability and data efficiency. At test time the tool properties, desired trajectory, and desired application forces could differ substantially from training scenarios. Adapting to this necessitates more data or computationally expensive online policy updates. In this work, we apply the principles of discrete dynamic programming (DP) to augment RL performance with domain knowledge. Specifically, we first design a computationally simple approximation of our environment. We then demonstrate in physical simulation that performing tree searches (i.e., lookaheads) and policy rollouts with this approximation can improve an RL-derived grasp sequence policy with minimal additional online computation. Additionally, we show that pretraining a deep RL network with the DP-derived solution to the discretized problem can speed up policy training.

Despite remarkable successes in solving various complex decision-making tasks, training an imitation learning (IL) algorithm with deep neural networks (DNNs) suffers from the high computation burden. In this work, we propose quantum imitation learning (QIL) with a hope to utilize quantum advantage to speed up IL. Concretely, we develop two QIL algorithms, quantum behavioural cloning (Q-BC) and quantum generative adversarial imitation learning (Q-GAIL). Q-BC is trained with a negative log-likelihood loss in an off-line manner that suits extensive expert data cases, whereas Q-GAIL works in an inverse reinforcement learning scheme, which is on-line and on-policy that is suitable for limited expert data cases. For both QIL algorithms, we adopt variational quantum circuits (VQCs) in place of DNNs for representing policies, which are modified with data re-uploading and scaling parameters to enhance the expressivity. We first encode classical data into quantum states as inputs, then perform VQCs, and finally measure quantum outputs to obtain control signals of agents. Experiment results demonstrate that both Q-BC and Q-GAIL can achieve comparable performance compared to classical counterparts, with the potential of quantum speed-up. To our knowledge, we are the first to propose the concept of QIL and conduct pilot studies, which paves the way for the quantum era.

Booking control problems are sequential decision-making problems that occur in the domain of revenue management. More precisely, freight booking control focuses on the problem of deciding to accept or reject bookings: given a limited capacity, accept a booking request or reject it to reserve capacity for future bookings with potentially higher revenue. This problem can be formulated as a finite-horizon stochastic dynamic program, where accepting a set of requests results in a profit at the end of the booking period that depends on the cost of fulfilling the accepted bookings. For many freight applications, the cost of fulfilling requests is obtained by solving an operational decision-making problem, which often requires the solutions to mixed-integer linear programs. Routinely solving such operational problems when deploying reinforcement learning algorithms may be too time consuming. The majority of booking control policies are obtained by solving problem-specific mathematical programming relaxations that are often non-trivial to generalize to new problems and, in some cases, provide quite crude approximations. In this work, we propose a two-phase approach: we first train a supervised learning model to predict the objective of the operational problem, and then we deploy the model within reinforcement learning algorithms to compute control policies. This approach is general: it can be used every time the objective function of the end-of-horizon operational problem can be predicted, and it is particularly suitable to those cases where such problems are computationally hard. Furthermore, it allows one to leverage the recent advances in reinforcement learning as routinely solving the operational problem is replaced with a single prediction. Our methodology is evaluated on two booking control problems in the literature, namely, distributional logistics and airline cargo management.

We propose a novel, object-agnostic method for learning a universal policy for dexterous object grasping from realistic point cloud observations and proprioceptive information under a table-top setting, namely UniDexGrasp++. To address the challenge of learning the vision-based policy across thousands of object instances, we propose Geometry-aware Curriculum Learning (GeoCurriculum) and Geometry-aware iterative Generalist-Specialist Learning (GiGSL) which leverage the geometry feature of the task and significantly improve the generalizability. With our proposed techniques, our final policy shows universal dexterous grasping on thousands of object instances with 85.4% and 78.2% success rate on the train set and test set which outperforms the state-of-the-art baseline UniDexGrasp by 11.7% and 11.3%, respectively.

We study generalizable policy learning from demonstrations for complex low-level control tasks (e.g., contact-rich object manipulations). We propose an imitation learning method that incorporates the idea of temporal abstraction and the planning capabilities from Hierarchical RL (HRL) in a novel and effective manner. As a step towards decision foundation models, our design can utilize scalable, albeit highly sub-optimal, demonstrations. Specifically, we find certain short subsequences of the demos, i.e. the chain-of-thought (CoT), reflect their hierarchical structures by marking the completion of subgoals in the tasks. Our model learns to dynamically predict the entire CoT as coherent and structured long-term action guidance and consistently outperforms typical two-stage subgoal-conditioned policies. On the other hand, such CoT facilitates generalizable policy learning as they exemplify the decision patterns shared among demos (even those with heavy noises and randomness). Our method, Chain-of-Thought Predictive Control (CoTPC), significantly outperforms existing ones on challenging low-level manipulation tasks from scalable yet highly sub-optimal demos.