AI and Machine Learning for Autonomous Systems

Vision for the Cluster

A heuristic definition of AI is the development of agents that perceive the world and use these observations to learn about the world, plan and reason. This is tightly connected to Machine Learning, where agents are given the ability to learn from data examples rather than being explicitly programmed. Common to projects in this cluster is that AI and Machine Learning are used in different ways to make autonomous systems smarter and more capable.

The objective of the cluster is to provide a platform for exchanging ideas and experiences about AI and Machine Learning methodology development in Autonomous Systems and Software. The activities in the cluster will be coordinated with the new WASP AI initiative – more information about this in due course!

The AI and Machine Learning cluster is intimately connected to the other WASP clusters, as it is methodology oriented – the AI and Machine Learning approaches developed within the cluster are applied in different areas of Autonomous Systems and Software. This is reflected by the large share of WASP PhD students being associated with the cluster — 21 as a primary affiliation (see below) and another 32 as secondary affiliation. This amounts to more than half of the total number of students within WASP.

Research and Industrial Challenges

The overlapping fields of AI and Machine Learning have grown enormously during the last decade. This growth is fuelled by three factors: 1) recent advances in computational hardware design that make it possible to implement large computational models, 2) the rapidly increasing volumes of data available through the Internet and different kinds of sensors, and 3) significant advances in methodologies for inference (i.e., optimization methods to fit the computational models to data). These three factors have together matured enough to make it possible to build AI and Machine Learning methods that work in realistic applications.

The recent growth in importance of AI and Machine Learning for the Swedish industry is clearly reflected in the WASP project with the creation of the new extension WASP AI. For further information, see the WASP AI homepage.

Sub-projects

 

Batch 1 (industrial

 

Vidit Saxena, KTH-Ericsson

Supervisor: Joakim Jaldén

Discriminative Learning & Optimization in Radio Access Networks

Using machine learning to model and optimize over wireless radio access networks.

 

Erik Jakobsson, LiU-Epiroc (prev AtlasCopco)

Supervisor: Erik Frisk

Predictive Maintenance for Autonomous Mining

How to enable more robust continuous operation of autonomous mining vehicles.

 

Batch 1 (university)

 

Olov Andersson, LiU

Supervisor: Patrick Doherty

Symbiotic Human-Robotic Interaction and Collaborative Planning in Distributed Knowledge-Rich Environments

Enabling robots and humans to collaborate, interact, and plan together, using different kinds of reasoning systems.

 

Mattias Tiger, LiU (affiliated)

Supervisor: Fredrik Heintz

Online Learning and Stream Reasoning for Situation Awareness in Robotics

Unsupervised learning of probabilistic spatio-temporal models of the robot itself and its environment, and incremental logical reasoning over these models.

 

Batch 2 (industrial

 

Hannes Eriksson, Chalmers-VolvoCars

Supervisor: Christos Dimitrakakis

Safe reinforcement learning agents for vehicular navigation and control

Develop autonomous and semi-autonomous AI systems embedded in vehicles and navigational systems that collaborate and interact with each other and with humans, using reinforcement learning.

 

Jens Henriksson, Chalmers-Semcon

Supervisor: Christian Berger

Machine Learning Multi Goal Model Optimization

Develop ML-based ptimizarion methods for real-time autonomous industrial applications. Parameters to optimize include physical energy consumption, execution time and computational requirement.

 

Erik Lindén, KTH-Tobii

Supervisor: Alexandre Proutière

Real-time Machine Learning for Embedded Devices: Low Complexity, Low Memory Sketches of Neural Networks

Developing NN models that are as simple as possible for the representation task at hand, in order to operate on hardware for gaze tracking, with limited computational resources.

 

Martin Isaksson, KTH-Ericsson

Supervisor: Seif Haridi

Deployment of Machine Intelligence functions for automation in distributed RAN environments

Techniques for incremental learning from streaming data with very limited computational resources, in cloud-based distributed Radio Access Networks.

 

Caroline Svahn, LiU-Ericsson

Supervisor: Mattias Villani

Machine learning for 5G System Control and Automation

Employing probabilistic machine learning models to increase the autonomy of 5G networks with the goal to improve network performance, reduce operation complexity, and increase resilience.

 

Olivier Moliner, LU-SonyMobile

Supervisor: Kalle Åström

Distributed Online Learning for Constrained IoT Devices

Developing systems of self-configured and nearly autonomous IoT-devices, with long battery life and low cost,  that autonomously detect and adapt to changes in their environment, study techniques to cooperatively spread relevant knowledge within the system.

 

Johan Källström, LiU-SAAB (affiliated)

Supervisor: Fredrik Heintz

LVC Simulation for Improved Training Efficiency

Machine Learning techniques for generation of intelligent behaviors for Computer Generated Forces (CGF) intended to support Live, Virtual and Constructive (LVC) simulation. The project will improve pilot training by offering more attractive LVC training scenarios and, as a result, improve the efficiency, effectiveness and readiness of the air force.

 

Batch 2 (university

 

Mina Ferizbegovic, KTH

Supervisor: Håkan Hjalmarsson

Learning Dynamical Systems

Learning dynamical systems relates to machine learning, cyber-physical systems and real-time big data analytics . This project covers 1) fundamental techniques for learning models in a statistical and computationally efficient way, 2) active and on-line learning for such models, 3) dynamical networked systems, relevant e.g. for 5G communication technology.

 

Samuel Murray, KTH

Supervisor: Hedvig Kjellström

Principled Integration of Logic Reasoning and Deep Learning

Deep Neural Networks (DNN) are extremely successful but require large training data. Humans are able to learn complex models from very few examples with logic reasoning. We want to combine DNN with logic reasoning and probabilistic models in a principled manner to achieve human-like learning.

 

Peter Varnai, KTH

Supervisor: Dimos Dimarogonas

Reinforcement learning endowed multi-robot planning and control under temporal logic tasks

Develop methods for distributed task and motion planning for multi-robot systems in unknown and dynamic environments, using distributed reinforcement learning. The research will blend elements from machine learning, distributed control and formal verification.

 

Fredrik Präntare, LiU

Supervisor: Fredrik Heintz

Probabilistic-Logic Stream Reasoning and Learning for Safe Autonomous Systems

Develop incremental techniques for learning and reasoning with statistical-relational models that combine the power of first-order logic and probability theory to represent, learn and reason with complex and uncertain relational structures in a principled manner. The system should be continually improving through semi-supervised on-line learning.

 

Hector Rodriguez-Deniz, LiU

Supervisor: Mattias Villani

Bayesian Learning for Spatio-Temporal Processes

Develop probabilistic models for spatio-temporal data with Bayesian learning methods, e.g. multi-output Gaussian processes and Gaussian random fields. Emphasis will be given to network data.

 

Erik Gärtner, LU

Supervisor: Cristian Sminchisescu

Deep People: Learning Integrated Visual Human Sensing Models

Develop visual human sensing methods based on large scale deep learning techniques. Models integrate person localization, pose estimation, and action and intent recognition based on images and video data with applications in robotics, entertainment and virtual reality.

 

Christopher Blöcker, UmU

Supervisor: Martin Rosvall

Self-driving transparent analytics
Develop transparent relational machine learning algorithms that can reveal patterns and causal mechanisms in prediction and decision problems. With better understanding of the inner-workings of machine learning algorithms, we can enable more efficient analysis, verification, and trusted decisions.

 

Timotheus Kampik, UmU

Supervisor: Helena Lindgren

Socially Intelligent Systems for Human-Agent Collaboration

Develop socially intelligent software agents that can collaborate with humans in decision making tasks to achieve goals and make prioritizations among potentially conflicting goals, needs, motivations, preferences and choices of actions.

Important aspects transparency and grounding in theories about human behaviour.

 

Lissy Pellaco, KTH (affiliated)

Supervisor: Joakim Jaldén

Actively Enhanced Cognition based Framework for Design of Complex Systems (AGNOSTIC)

The project will consider the merger of traditional parameterized and data-driven approaches in the context of radio access networks, with a focus towards on active cognition.

 

Xuechun Xu, KTH (affiliated)

Supervisor: Joakim Jaldén

Automating System SpEcific Model-Based LEarning (ASSEMBLE)

The project considers Bayesian inference techniques, discriminative and generative, for machine learning with a focus on modularity and implementation using probabilistic programming paradigms.