Sample-Efficient Curriculum Reinforcement Learning for Complex Reward Functions

Reinforcement learning (RL) shows promise in control problems, but its practical application is often hindered by the complexity arising from intricate reward functions with constraints. While the reward hypothesis suggests these competing demands can be encapsulated in a single scalar reward function, designing such functions remains challenging. Building on existing work, we start by formulating preferences over trajectories to derive a realistic reward function that balances goal achievement with constraint satisfaction in the application of mobile robotics with dynamic obstacles. To mitigate reward exploitation in such complex settings, we propose a novel two-stage reward curriculum combined with a flexible replay buffer that adaptively samples experiences. Our approach first learns on a subset of rewards before transitioning to the full reward, allowing the agent to learn trade-offs between objectives and constraints. After transitioning to a new stage, our method continues to make use of past experiences by updating their rewards for sample-efficient learning. We investigate the efficacy of our approach in robot navigation tasks and demonstrate superior performance compared to baselines in terms of true reward achievement and task completion, underlining its effectiveness.

ECAP: Extensive Cut-and-Paste Augmentation for Unsupervised Domain Adaptive Semantic Segmentation

We consider unsupervised domain adaptation (UDA) for semantic segmentation in which the model is trained on a labeled source dataset and adapted to an unlabeled target dataset. Unfortunately, current self-training methods are susceptible to misclassified pseudo-labels resulting from erroneous predictions. Since certain classes are typically associated with less reliable predictions in UDA, reducing the impact of such pseudo-labels without skewing the training towards some classes is notoriously difficult. To this end, we propose an extensive cut-and-paste strategy (ECAP) to leverage reliable pseudo-labels through data augmentation. Specifically, ECAP maintains a memory bank of pseudo-labeled target samples throughout training and cut-and-pastes the most confident ones onto the current training batch. We implement ECAP on top of the recent method MIC and boost its performance on two synthetic-to-real domain adaptation benchmarks. Notably, MIC+ECAP reaches an unprecedented performance of 69.1 mIoU on the Synthia->Cityscapes benchmark. Our code is available at https://github.com/ErikBrorsson/ECAP.

Falsification of Cyber-Physical Systems using Bayesian Optimization

Cyber-physical systems (CPSs) are usually complex and safety-critical; hence, it is difficult and important to guarantee that the system's requirements, i.e., specifications, are fulfilled. Simulation-based falsification of CPSs is a practical testing method that can be used to raise confidence in the correctness of the system by only requiring that the system under test can be simulated. As each simulation is typically computationally intensive, an important step is to reduce the number of simulations needed to falsify a specification. We study Bayesian optimization (BO), a sample-efficient method that learns a surrogate model that describes the relationship between the parametrization of possible input signals and the evaluation of the specification. In this paper, we improve the falsification using BO by; first adopting two prominent BO methods, one fits local surrogate models, and the other exploits the user's prior knowledge. Secondly, the formulation of acquisition functions for falsification is addressed in this paper. Benchmark evaluation shows significant improvements in using local surrogate models of BO for falsifying benchmark examples that were previously hard to falsify. Using prior knowledge in the falsification process is shown to be particularly important when the simulation budget is limited. For some of the benchmark problems, the choice of acquisition function clearly affects the number of simulations needed for successful falsification.

An SMT Based Compositional Algorithm to Solve a Conflict-Free Electric Vehicle Routing Problem

The Vehicle Routing Problem (VRP) is the combinatorial optimization problem of designing routes for vehicles to visit customers in such a fashion that a cost function, typically the number of vehicles, or the total travelled distance is minimized. The problem finds applications in industrial scenarios, for example where Automated Guided Vehicles run through the plant to deliver components from the warehouse. This specific problem, henceforth called the Electric Conflict-Free Vehicle Routing Problem (CF-EVRP), involves constraints such as limited operating range of the vehicles, time windows on the delivery to the customers, and limited capacity on the number of vehicles the road segments can accommodate at the same time. Such a complex system results in a large model that cannot easily be solved to optimality in reasonable time. We therefore developed a compositional model that breaks down the problem into smaller and simpler sub-problems and provides sub-optimal, feasible solutions to the original problem. The algorithm exploits the strengths of SMT solvers, which proved in our previous work to be an efficient approach to deal with scheduling problems. Compared to a monolithic model for the CF-EVRP, written in the SMT standard language and solved using a state-of-the-art SMT solver the compositional model was found to be significantly faster.