Mastering the game of Go with deep neural networks and tree search

Authors: David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, Demis Hassabis

For class EE/CSC 7700 ML for CPS

Instructor: Dr. Xugui Zhou

Presentation by Group 7: Rishab Meka (Presenter), Bharath Kollanur

Summarized by Group 6: Cheng Chen, Zhiyong Sui

Summary

This paper introduces a novel sequence transduction model architecture named the Transformer. This architecture is based solely on attention mechanisms, eliminating the need for recursion and convolution. The model addresses the limitations of sequence models that rely on recursive processes, which perform poorly in parallelization and computational efficiency for longer sequences. The Transformer adopts an encoder-decoder structure, where the encoder consists of identical layers with multi-head self-attention and fully connected feed-forward networks, while the decoder mirrors this structure but adds a multi-head attention layer on the encoder's output; utilizing scaled dot-product attention and multi-head attention, the model computes the importance of key-value pairs based on queries and allows joint attention across different subspaces, with encoder-decoder attention enabling the decoder to focus on all input positions, self-attention improving contextual understanding by attending to all positions within layers, and positional encodings ensuring the model captures the order of tokens in a sequence.


Slide Outlines

Motivation

History Image

Go poses a unique challenge for AI because of its vast search space, which is significantly larger than that of chess, making traditional AI techniques ineffective. Mastering Go is seen as a key test of AI’s ability to handle complex, strategic decision-making and long-term planning.

Introduction

Introduction Image Introduction Image

AlphaGo combines deep neural networks with Monte Carlo Tree Search (MCTS) to address the complexity of Go. It leverages two types of networks: policy networks to choose moves and value networks to evaluate board positions.

AlphaGo initially learns from human expert games (supervised learning) and then improves through self-play (reinforcement learning), achieving a near-perfect win rate against other Go programs and defeating a human European champion in a 5-0 series.

Architecture

Solution Image

The AlphaGo architecture integrates neural networks with MCTS. This architecture enables the system to learn complex patterns on the Go board and make strategic decisions based on past experiences, balancing both exploration and exploitation.

Policy Network

Residual Example 1

The policy network predicts the probability of possible moves in a given board state, using deep convolutional neural networks designed to recognize patterns specific to Go. This network helps AlphaGo prioritize moves based on learned strategies.

Supervised Learning Policy Network

Residual Example 2

AlphaGo’s initial policy network is trained through supervised learning, where it learns from 30 million game positions from the KGS Go Server. This network outputs a probability distribution over all legal moves, helping it replicate human-level play initially.

Reinforcement Learning Policy Network

Residual Blocks Image

AlphaGo enhances the policy network through reinforcement learning by playing games against past versions of itself. Randomly selecting previous versions as opponents prevents overfitting, ensuring the network continues to improve without bias towards recent strategies.

Value Network

Residual Blocks Image Residual Blocks Image

The value network evaluates board positions to estimate the likelihood of winning from a given state, outputting a single prediction rather than a move probability distribution. It is trained by minimizing prediction errors using stochastic gradient descent, which helps stabilize evaluation without overfitting.

Monte Carlo Tree Search (MCTS)

MCTS Process: MCTS iteratively explores possible moves by simulating actions and outcomes to select the best move. At each step, an action is chosen to maximize the action value plus an exploration bonus.

Evaluation Methods: Leaf nodes are evaluated both by the value network and through random rollouts (simulated games to completion), which improves the accuracy of action values for each subtree, allowing AlphaGo to refine its move choices further.

Results

Limitations Image Limitations Image

AlphaGo demonstrated strong general performance, achieving high success rates across different test scenarios. The combination of neural networks with Monte Carlo Tree Search allowed AlphaGo to perform at a level that surpassed all prior Go-playing AI models.

Results against other models

Aftermath Image

AlphaGo was tested against other leading Go models, consistently outperforming them and setting a new standard for AI capabilities in the game. Its success showcased the superiority of its deep learning and tree search integration over previous models, which relied more on simpler heuristic-based approaches.

Results against champion

Limitations Image

AlphaGo's ultimate test was a match against a human European Go champion, where it achieved a decisive 5-0 victory. This milestone highlighted AlphaGo’s ability to not only compete with but surpass human expertise, marking a significant achievement in AI game-playing and strategic thinking.

Conclusion

Aftermath Image

AlphaGo demonstrates that AI can achieve superhuman performance in complex tasks through self-play and reinforcement learning. The integration of deep learning with tree search techniques represents a significant advancement in AI’s potential for strategic reasoning.

Teamwork

Limitations Image

Paper’s References

Aftermath Image Aftermath Image


Discussion Questions

Discussion 1: What other domains other than game simulation can AlphaGo be used?

Financial Markets, Bookmakers

Discussion 2: What do you think are the limitations of this two-stage training approach, and how might this influence AlphaGo's capacity to discover unique strategies?

The limitation of two-stage training methods is that because they rely on supervised learning, they need some expert games to have a good starting point. And then in areas where you don't have access to those games, if you're just trying to use it for something new, there's no good strategy to start with.

This case is highly dependent on whether you can label, label the states, label the states correctly. So you can actually do learning supervision without this. The only difference in this approach is that it is able to learn the value function of some states. So in any real life, uncertain scenario, this is not a gap. I believe it is difficult to label these.