A5052941553- AI-based Problem Solving and Planning
- Machine Learning and Algorithms
- Constraint Satisfaction and Optimization
- Reservoir Engineering and Simulation Methods
- Multimodal Machine Learning Applications
- Explainable Artificial Intelligence (XAI)
- Oil and Gas Production Techniques
- Artificial Intelligence in Games
- Machine Learning and Data Classification
- Adversarial Robustness in Machine Learning
- Complex Systems and Decision Making
- Fatty Acid Research and Health
- Distributed systems and fault tolerance
- Ruminant Nutrition and Digestive Physiology
- Parallel Computing and Optimization Techniques
- Logic, Reasoning, and Knowledge
- Logic, programming, and type systems
- Intelligent Tutoring Systems and Adaptive Learning
- Advanced Graph Neural Networks
- Distributed and Parallel Computing Systems
- Reinforcement Learning in Robotics
- Robotic Path Planning Algorithms
- Teaching and Learning Programming
- Meat and Animal Product Quality
- Semantic Web and Ontologies
University of Basel
2018-2024
Saarland University
2022-2024
As classical planning is known to be computationally hard, no single planner expected work well across many domains. One solution this problem use online portfolio planners that select a for given task. These portfolios perform classification task, well-known and wellresearched task in the field of machine learning. The usually performed using representation tasks with collection hand-crafted statistical features. Recent techniques learning are based on automatic extraction features have not...
Automated planning is one of the foundational areas AI. Since no single planner can work well for all tasks and domains, portfolio-based techniques have become increasingly popular in recent years. In particular, deep learning emerges as a promising methodology online selection. Owing to development structural graph representations tasks, we propose neural network (GNN) approach selecting candidate planners. GNNs are advantageous over straightforward alternative, convolutional networks, that...
How can we train neural network (NN) heuristic functions for classical planning, using only states as the NN input? Prior work addressed this question by (a) per-instance imitation learning and/or (b) per-domain learning. The former limits approach to instances small enough training data generation, latter domains where necessary knowledge generalizes across instances. Here explore three methods that make generation scalable through bootstrapping and approximate value iteration. In...
Testing is a promising way to gain trust in neural action policies π. Previous work on policy testing sequential decision making targeted environment behavior leading failure conditions. But if the unavoidable given that behavior, then π not actually blame. For situation qualify as "bug" π, there must be an alternative π' does better. We introduce generic framework based intuition. This raises bug confirmation problem, deciding whether or state bug. analyze use of optimistic and pessimistic...
We propose a new approach based on ranking to learn guide Greedy Best-First Search (GBFS). As previous approaches, ours is the observation that directly learning heuristic function overly restrictive, and GBFS capable of efficiently finding good plans for much more flexible class total quasi-orders over states. In order an optimal function, we introduce framework leveraging any neural network regression model handling training data through batching. Compared with approaches planning, does...
Benchmark data sets are an indispensable ingredient of the evaluation graph-based machine learning methods. We release a new set, compiled from International Planning Competitions (IPC), for benchmarking graph classification, regression, and related tasks. Apart construction (based on AI planning problems) that is interesting in its own right, set possesses distinctly different characteristics popularly used benchmarks. The named IPC, consists two self-contained versions, grounded lifted,...
One of the strongest approaches for optimal classical planning is A* search with heuristics based on abstractions task. Abstraction are well studied in formalisms without conditional effects such as SAS+. However, crucial to model many tasks compactly. In this paper, we focus *factored* which allow a specific form effect, where variable x can only depend value x. We generalize projections, domain abstractions, Cartesian and counterexample-guided abstraction refinement method formalism. While...
Neural networks (NN) are increasingly investigated in AI Planning, and used successfully to learn heuristic functions. NNs commonly not only predict a value, but also output confidence this prediction. From the perspective of search with NN heuristics, it is natural idea take into account, e.g. falling back standard where low. We contribute an empirical study idea. design methods which prune nodes, or switch between queues, based on NNs. furthermore explore possibility out-of-distribution...
Automated planning is one of the foundational areas AI. Since no single planner can work well for all tasks and domains, portfolio-based techniques have become increasingly popular in recent years. In particular, deep learning emerges as a promising methodology online selection. Owing to development structural graph representations tasks, we propose neural network (GNN) approach selecting candidate planners. GNNs are advantageous over straightforward alternative, convolutional networks, that...
Previous work introduced the concept of progress states. After expanding a state, greedy best-first search (GBFS) will only expand states with lower heuristic values. Current methods can identify for single task and after solution has been found. We introduce novel approach that learns description logic formula characterizing all in classical planning domain. Using learned formulas GBFS to break ties favor often significantly reduces effort.
Since no classical planner consistently outperforms all others, it is important to select a that works well for given planning task. The two strongest approaches selection use image and graph convolutional neural networks. They have the drawback learned models are complicated uninterpretable. To obtain explainable models, we identify small set of simple task features show elementary interpretable machine learning techniques can these solve roughly as many tasks complex based on