The Role of Multi-Agent Learning in Artificial Intelligence Research at DeepMind

Explore the role of multi-agent learning in artificial intelligence research at DeepMind in this comprehensive lecture from the Alan Turing Institute. Delve into the concept of intelligence as an agent's ability to achieve goals in diverse environments, with a focus on evolving collections of agents. Examine two key projects: the study of cooperation among self-interested agents using Sequential Social Dilemmas, and the groundbreaking AlphaGo project that utilized Learning from Self-Play to defeat top professional Go players. Gain insights into the challenges and advancements in multi-agent learning, including temporal dynamics, coordination problems, and the complexities of the game of Go. Discover the innovative approaches used in AlphaGo, such as value networks, policy networks, and supervised learning techniques. Analyze the lessons learned from these projects and their implications for the future of AI research.

Introduction
Welcome
About DeepMind
What is Intelligence
Multiagent Systems
Multiagent Aspects
Cumulative Culture
Social Dilemmas
Results
Conclusion
The Game of Go
Why is Go so difficult
Game Space Complexity
Value Network
Policy Network
Human Expert Game Records
Supervised Policy Network
Train Value Network
Supervised Learning
Value Networks
Evaluation
Random Roll
Evaluation of Go
Innovation in Go
Alphago test games
Alphago team
Lessons from Alphago
What hasnt been achieved