Machine Intelligence

Featured collaboration

University of Oxford (opens in new tab)

MSR PI: Katja Hofmann (opens in new tab)
University of Oxford PI: Shimon Whiteson (opens in new tab)
Joint Postdoctoral Researcher: Mingfei Sun

Reinforcement Learning for Gaming

This project will focus on developing and analysing state-of-the-art reinforcement learning (RL) methods for application to video games.  The project aims to tackle two key challenges.  First, building effective game AI with RL requires dramatically scaling up existing tools for cooperative multi-agent RL, in which teams of agents must collaborate to complete tasks.  Doing so requires new methods for performing multi-agent credit assignment and multi-agent exploration in large state and action spaces.  Second, effective game AI must also be able to transfer effectively to new scenarios, such as new game levels and versions, without having to learn from scratch.  Doing so requires new methods for transfer and meta-learning in RL that scale to the complexity of modern video games.

Industry collaborators

Ninja Theory (opens in new tab)

Ninja Theory was formed in 2004 by four partners, including current Directors Nina Kristensen (Chief Development Director), Tameem Antoniades (Chief Creative Director) and Jez San OBE (Non-Executive Director). The studio pride themselves on striving for the highest production values and continually pushing the boundaries of technology, art and design to create evermore exciting video game experiences.

Find out more about our collaboration with Ninja Theory on the Project Paidia page >

IGGI Centre for Doctoral Training (opens in new tab)

Industry Partner and Advisory Board Member of the IGGI Centre for Doctoral Training. (opens in new tab)

Academic Collaborations

Berkely AI Research

Learning to Collaborate with Human Players
Katja Hofmann (opens in new tab) (MSR Cambridge), Sam Devlin (opens in new tab) (MSR Cambridge), Kamil Ciosek (opens in new tab) (MSR Cambridge), Professor Anca Dragan (opens in new tab) (BAIR), Micah Carroll (PhD student)

Find out more on our Berkeley AI Research collaboration page >

Queen Mary University London (opens in new tab)

Malmo 2020 Multi-Agent Upgrade

Diego Perez Liebana (opens in new tab)

Microsoft’s Project Malmo platform enables users to create worlds and learning agents able to play multiple 3D games within Minecraft. In recent years, we have co-organised two international competitions. First on multi-agent learning and, secondly, on sample efficient reinforcement learning with human priors . These competitions have extended the features of the platform, but each introduced their own API, installation instructions and documentation, which has created an unnecessary barrier to researchers wanting to get started with the platform. The objective of this project is to unify the extensions from both competitions back into the original Malmo benchmark, to provide a common entry point for researchers.

PhD collaborations in EMEA

• Max Planck Institute for Software Systems

  Reinforcement Learning for Enabling Next Generation Human-Machine Partnerships

  MSR Supervisor: Sam Devlin

  External Supervisor: Adish Singla (opens in new tab)

• Queen Mary University

  Local Forward Model Learning for Sample-Efficient Sequential Decision Making in Open-World 3D Games

  MSR Supervisor: Sam Devlin

  External Supervisor: Diego Perez Liebana (opens in new tab)

• University of York

  Deep Reinforcement Learning For Collaborative Game AI To Enhance Player Experience

  MSR Supervisor: Sam Devlin

  External Supervisor: TBC

• University of Edinburgh

  Better Sample Efficiency of Reinforcement Learning

  MSR Supervisor: Kamil Ciosek

  External Supervisor: Amos Storkey (opens in new tab)

• University of Oxford

  Reinforcement Learning for Adaptive User Interaction

  MSR Supervisor: Katja Hofmann

  External Supervisor: Shimon Whiteson (opens in new tab)

• INRIA

  Intrinsically Motivated Exploration for Lifelong Deep Reinforcement Learning of Multiple Tasks

  MSR Supervisor: Katja Hofmann

  External Supervisor: Pierre-Yves Oudeyer (opens in new tab)