David Silver

Developed a reinforcement learning algorithm that achieves superhuman proficiency in the game of Go starting from random play without human data, guidance, or domain knowledge beyond basic rules.

Achieved a 100-0 win record against the champion-defeating version of AlphaGo.

Replaced separate policy and value networks with a single neural network architecture consisting of many residual blocks of convolutional layers with batch normalization and rectifier non-linearities.

Implemented a simplified tree search that relies on a single neural network for position evaluation and move sampling without performing Monte-Carlo rollouts.

Incorporated lookahead search inside the training loop to achieve rapid improvement and stable learning.

Rediscovered fundamental elements of human Go knowledge, including joseki (corner sequences), fuseki (opening strategies), and life-and-death concepts from first principles.

Optimized the system to run on a single machine with 4 Tensor Processing Units (TPUs) in the Google Cloud.

Defeated Lee Sedol, the winner of 18 international titles, in March 2016.

Implemented a distributed system over multiple machines utilizing 48 Tensor Processing Units (TPUs) to evaluate neural networks during search.

Trained the value network using outcomes from fast games of self-play by AlphaGo.

Employed a larger policy and value network architecture compared to earlier versions, featuring 12 convolutional layers.