Contents

## How does AlphaGo Zero work?

In AlphaGo Zero, we use a single deep network f, composed of convolutional layers, to estimate both p and v. It takes the board position (s) as input and outputs p and v accordingly. In the original AlphaGo, it uses two separate deep networks, and we train them with human-played games using supervised learning.

## What are the differences between AlphaGo Zero and its predecessors how did these differences improve AlphaGo Zero’s performance?

AlphaGo Zero’s strategies were self-taught i.e it was trained without any data from human games. AlphaGo Zero was able to defeat its predecessor in only three days time with lesser processing power than AlphaGo. However, the original AlphaGo, on the other hand required months to learn how to play.

**Can a human beat AlphaGo?**

Lee Se-dol is the only human to ever beat the AlphaGo software developed by Google’s sister company Deepmind. In 2016, he took part in a five-match showdown against AlphaGo, losing four times but beating the computer once. AlphaGo was developed by Deepmind, which is owned by Google’s parent company Alphabet.

**How is the expert policy represented in AlphaGo Zero?**

The expert policy and the approximate value function are both represented by deep neural networks. In fact, to increase efficiency, Alpha Zero uses one neural network that takes in the game state and produces both the probabilities over the next move and the approximate state value.

### How is a leaf selected in AlphaGo Zero?

Instead, the selection process chooses nodes that strike a balance between being lucrative-having high estimated values-and being relatively unexplored-having low visit counts. A leaf node is selected by traversing down the tree from the root node, always choosing the child with the highest upper confidence tree (UCT) score:

### How does the alpha zero algorithm work over time?

The Alpha Zero algorithm produces better and better expert policies and value functions over time by playing games against itself with accelerated Monte Carlo tree search. The expert policy and the approximate value function are both represented by deep neural networks.

**How are values of close to zero improved?**

Values of close to zero produce policies that choose the best move according to the Monte Carlo tree search evaluation. The value portion of the neural network is improved by training the predicted value to match the eventual win/loss/tie result of the game, . Their loss function is: