Unveiling the Mastery of AlphaGo: Deep Neural Networks and Tree Search

Unveiling the Mastery of AlphaGo: Deep Neural Networks and Tree Search

Table of Contents:

  1. Introduction
  2. The Game of Go
  3. Challenges in Solving Go
  4. The Rise of AlphaGo
  5. Understanding AlphaGo's Approach
    • 5.1 Policy Network
    • 5.2 Rollout Policy
    • 5.3 Reinforcement Learning Policy
    • 5.4 Value Network
  6. Monte Carlo Tree Search
    • 6.1 UCB Algorithm
    • 6.2 Virtual Loss
    • 6.3 Tree Policy
  7. Symmetries in AlphaGo
  8. Computation and Training
  9. Conclusion
  10. Resources


In this article, we will delve into the fascinating world of AlphaGo, the groundbreaking AI program developed by DeepMind that became a sensation by defeating a world champion Go player. We will explore the game of Go, the challenges it poses, and the approach AlphaGo took to master it. This article aims to provide a comprehensive understanding of the technical aspects behind AlphaGo's success while explaining complex concepts in a simple and engaging manner. So, let's dive in and unravel the mysteries of AlphaGo!

The Game of Go

Before we delve into AlphaGo, let's first understand the game that captivated the AI community - Go. Go is an ancient board game that originated in China over 2,000 years ago. It is played on a 19x19 GRID, with players placing white and black stones on the intersections of the lines. The objective of the game is to acquire as much territory as possible while strategically capturing the opponent's stones. Due to its simple rules and complex gameplay, Go has been a significant challenge for AI algorithms.

Challenges in Solving Go

Unlike other board games like chess or checkers, Go has a much larger search space. The number of possible configurations on the Go board is more than the number of atoms in the Universe. This complexity makes it highly challenging for traditional AI algorithms, such as minimax search or depth minimax search, to solve the game. Previous approaches, including pure Monte Carlo research methods, failed to provide effective solutions for Go. The daunting task of solving Go seemed out of reach until AlphaGo emerged.

The Rise of AlphaGo

AlphaGo burst onto the scene in 2015-2016, revolutionizing the world of AI. Developed by DeepMind, AlphaGo cracked the game of Go, achieving what was once considered impossible. It marked a significant breakthrough in the AI community and sparked a widespread interest in the potential of deep learning algorithms. The success of AlphaGo paved the way for advancements in various other domains that were previously believed to be insurmountable.

Understanding AlphaGo's Approach

To comprehend how AlphaGo conquered the game of Go, we need to understand the different components that make up its algorithm. AlphaGo's approach can be divided into four main parts: the policy network, the rollout policy, the reinforcement learning policy, and the value network. Let's explore each of these components in detail.

Policy Network

The policy network plays a crucial role in AlphaGo's decision-making process. It takes the current state of the game as input, represented by a 19x19 feature map, and outputs a probability distribution over all possible moves. The policy network is trained using Supervised learning on a vast database of expert Go games. By learning from expert moves, the policy network can predict the most promising moves to play in a given state. This initial policy network provides a strong foundation for subsequent improvements.

Rollout Policy

The rollout policy is a Simplified version of the policy network that uses a linear softmax model. It is much smaller and faster than the policy network, making it suitable for quick evaluations during the Monte Carlo tree search process. The rollout policy serves as a fast approximation of the game's outcome, providing a basis for exploration during simulations.

Reinforcement Learning Policy

The reinforcement learning policy is the heart of AlphaGo's learning mechanism. It is trained using a reinforcement learning approach called the REINFORCE algorithm. The RL policy is initialized with the pre-trained supervised learning policy network and further improved through a process of self-play and policy gradient methods. The RL policy learns from its own gameplay and continuously updates and refines its strategy to achieve better outcomes.

Value Network

The value network is responsible for estimating the value or win probability of a given state. It takes the current game state as input and outputs a scalar value indicating the likelihood of winning. The value network is trained using a modified version of the RL policy network, where the objective is to minimize the mean squared error between the predicted outcome and the actual game result. By predicting the game outcome accurately, the value network provides critical information for the Monte Carlo tree search process.

Monte Carlo Tree Search

To navigate through the vast search space of Go, AlphaGo employs a technique called Monte Carlo tree search (MCTS). MCTS is a heuristic search algorithm that uses simulated gameplay to guide its decision-making process. It gradually builds a search tree by exploring different paths and approximating the values of each node through repeated simulations. Let's delve into the key aspects of MCTS used by AlphaGo:

UCB Algorithm

AlphaGo uses a variant of the upper confidence bound (UCB) algorithm called PUCT (polynomial upper confidence bound for trees). UCB balances exploration and exploitation by considering both the action value function and a prior probability distribution. This balance allows AlphaGo to make informed decisions, favoring unexplored paths initially and transitioning towards the most promising paths as it gathers more information.

Virtual Loss

To improve concurrency and discourage duplicate explorations, AlphaGo utilizes a technique called virtual loss. Virtual loss artificially increases the visitation count of an edge and temporarily reduces the estimated action values during the tree search. This approach ensures that other Threads do not duplicate the same exploration simultaneously, leading to more efficient and effective simulations.

Tree Policy

The tree policy guides the search process by selecting actions during simulations. It balances between exploration and exploitation based on the UCB algorithm. Initially, the tree policy prefers actions with high prior probabilities and low visit counts. However, as the search progresses, it asymptotically favors actions with high action values, ensuring more accurate and promising gameplay.

Symmetries in AlphaGo

To enhance the efficiency and effectiveness of AlphaGo's search, symmetries in the game of Go are exploited. By leveraging rotational and reflection symmetries, AlphaGo uses minimal network evaluations to Extrapolate the action probabilities for all symmetric positions. These symmetries allow AlphaGo to augment its data set and improve its predictions, making it more robust and versatile in gameplay.

Computation and Training

Training AlphaGo requires significant computational resources. In the case of the supervised learning (SL) policy network, training takes three weeks using 50 GPUs. The reinforcement learning (RL) policy network is trained for one day using the same number of GPUs. The value network, on the other HAND, is trained for a week using 50 GPUs. The immense computational power required underscores the complexity of the AlphaGo algorithm and its training process.


AlphaGo's journey from inception to becoming an unbeatable Go player demonstrated the remarkable potential of deep learning algorithms. By combining advanced techniques such as deep neural networks, reinforcement learning, and Monte Carlo tree search, AlphaGo achieved Superhuman capabilities in a game previously thought to be unsolvable. The lessons learned from AlphaGo's success have paved the way for solving complex AI problems in various domains and continue to propel advancements in the field.


  • DeepMind's AlphaGo Paper: Link

Note: The content of this article is sourced from DeepMind's AlphaGo research paper and is intended to provide a comprehensive overview of the topic. For more detailed information and technical specifications, refer to the original research paper.

Most people like

Find AI tools in Toolify

Join TOOLIFY to find the ai tools

Get started

Sign Up
App rating
AI Tools
Trusted Users
No complicated
No difficulty
Free forever
Browse More Content