Our Research

London Reinforcement Learning Team

haitham

Haitham Bou Ammar

Expert in various areas of machine learning and optimisation

rasul

Rasul Tutunov

Expert in convex and non-convex optimisation

mohammad

Mohammed Amin Abdullah

Expert in reinforcement learning and operations research

aivar

Aivar Sootla

Expert in control theory

victor

Victor Gabillon

Expert in derivative-free optimisation, bandits, online, and reinforcement learning

yaodong

Yaodong Yang

Expert in multi-agent systems and reinforcement learning

Huawei Reinforcement Learning Research

Reinforcement Learning Team in London

Aim & Direction: Our aim in the reinforcement learning team in London is to enable useful and effective decision making in ubiquitous applications including but not limited to, self-driving cars, logistics and telecommunication networks. We realise that current techniques for automated decision-making lack in:

To tackle these limitations and enable next-generation autonomous decision-makers, we have assembled a world-leading team of researchers and engineers with diverse expertise in probability, control, numerical optimisation, reinforcement learning, and game theory. To ensure effective execution, we organised the team in three coherent groups uniting to solve each of the above challenges:

Research Projects

Safe and Robust Decision-Makers

Though successful in well-behaving simulated environments (e.g., computer games), current reinforcement learning techniques are unsafe and tend to over-fit to training environments. When it comes to real-world applications, however, safety and robustness considerations are key to successful and autonomous decision-makers --a self-driving car, for instance, can not crash due to a random actions,or slight changes in driving conditions (e.g., weather conditions, modelling assumptions).

State-of-the-art Reinforcement Learning - Training System

Wasserstein Robust Reinforcement Learning - Training System

State-of-the-art Reinforcement Learning - Training System

Wasserstein Robust Reinforcement Learning - Training System

Strategic and Scalable Decision Makers

Another important component in decision-making for the real-world involves multiple interacting entities (multi-agent), with potentially conflicting goals, operating together in a shared environment. Current methods from multi-agent reinforcement learning suffer in scalability with results reported on a handful number of agents (e.g., three to four). Our goal is to scale strategic decision-makers to support large strategy spaces both during evaluation and learning.

Optimisation-Theoretic Decision Makers

At the core of any machine learning algorithm is a numerical optimisation algorithm attempting to determine a solution to minimising a loss function that encodes the designer’s goals, e.g., consider a convolutional network with a logistic loss while using ADAM. Similarly, reinforcement learning defines total-expected return objectives and optimises for effective policies. Contrary to static problems, however, available data in reinforcement learning is affected by updated parameters, i.e., trajectory density is parameterised by free variables. We believe substantial improvements to decision-making in real-world applications can come from developing better-behaved optimisers that consider stochasticity, and non-stationarity. To this end, we research in two directions:

Kirin AI

The future of AI is on-device as more and more functionality and processing capability consumer devices gaining over-time. Also, with privacy laws such as GDPR being stricter on the movement of user data, AI has to function independent of the cloud or the service providers. Currently, AI services can’t manage to provide reliable customer satisfaction without relying on the cloud except for very simple tasks.

Kirin AI focuses on developing both the chipset capabilities and optimized algorithms for AI to function optimal lyon-chip without any dependency on the cloud or the need to move user data off-device. With this objective comes a lot of challenges which Kirin AI is addressing in its research.

Our main focus is to push the forefront of AI technology in terms of performance. Then, research adapting our AI algorithms to best function on-device using our latest Kirin Chipset.

Kirin AI UK has teams in Huawei Cambridge and London Research centres. Our research is focused on text, speech and video modalities. In Speech, we are researching different domains in Automatic Speech Recognition. In Video, we are researching super resolution, 2D-3D pose estimation and tracking and action recognition. Our research is dominated by three key factors: performance, memory footprint and latency.

Kirin AI are interested in fresh talent who are eager to take on the challenge to push the frontiers of AI, Machine Learning and Deep Learning within application domains of NLP, Computer Vision and Speech. Our AI competition is based on the very essence of what the wider research community is working on. We aim on aligning both students and researches on the same objectives to excel in overcoming current AI obstacles.

Further information

To know more about our worldwide research, please visit Huawei Noah’s Ark Lab and Huawei Hisilicon.

Return to homepage