The game AlphaZero – Conquers Chess, Shogi, and Go has won the hearts of countless gamers around the world. In this article, we’ll discuss the Generalization of AlphaGo Zero, Examples of games played by AlphaZero against Stockfish, and an evaluation of AlphaZero’s performance in chess and go.
Table of Contents
AlphaZero – Conquers chess, shogi and go
AlphaZero uses a layered model to measure its performance against humans as well as against a set opening positions from the 2016 TCEC. AlphaZero beat Elmo in a head to head tournament by more than two percentage point. It also scored a perfect score for Go. Unlike other AI programs, AlphaZero has no deterministic strategy and instead bases its decisions on the current state of the game.
To make use of these techniques, AlphaZero uses a table S1 that contains data on each game position in the game. Each of these data features are repeated for each position in T = 8-step history. The tablebase contains counts, as well as various other features such as the typical number of legal moves. The tablebase also shows the current player and opponent, denoted respectively by P1 or P2.
AlphaZero, despite these differences, is a highly dynamic computer software program. It optimizes its own pieces’ activity while minimising those of its competitors. AlphaZero’s game play also seems to place less emphasis on “material”, as modern chess does. AlphaZero can still win while playing like a madman.
AlphaZero AI, an artificial intelligence (AI), program, can master any game in less that a week. The system was created by DeepMind, a subsidiary of Alphabet, which has the resources and expertise to create such a machine. The AlphaZero algorithm started life last year by beating an AI program that specialized in Go. This AI was programmed to play the game with the best moves. It took only three days for the AI to learn, making it a super-human in just hours.
Generalization of AlphaGo Zero
AlphaGo Zero, a computer program designed to play Go better than humans, is a generalization of AlphaGo Zero. Although the program can play chess, shogi and other games, the ultimate goal is to be a world-class Go player. AlphaGo Zero is based on the same approach as AlphaGo, but with a few major differences. This article discusses these differences, and explains how AlphaGo Zero compares to human players in each of these games.
The generalization of AlphaGo is not a reflection of how useful AI techniques are. Go is the most basic category of AI tasks. This game is easy to learn and has all the characteristics that you would expect from an easy-to-learn task. It is simple to play, easy to score, and cheap to simulate. The branching factor is the only thing that makes AlphaGo difficult. DeepMind researchers thoroughly analyzed the game to create their algorithm.
The software has been able, over the past few years to master the game better than human players. AlphaGo Zero begins reinforcement learning instead of training against humans. At first, it plays random moves and learns to play against itself. It beat all previous AlphaGo versions, including AlphaGo Lee. And, it can beat humans in as little as a day. The program is now capable of outsmarting human players in Go.
Example games played by AlphaZero against Stockfish
The context-dependent positional assessment, or CDE, is one of the most fascinating aspects in chess games. In the following example games played by AlphaZero against Stockfish, the computer sacrificed a few pieces to gain long-term strategic advantage. AlphaZero has an advantage over Stockfish in many areas including king-side positioning. However, its weakness lies in defensive strategy where its pieces are often sacrificed to gain positional advantages.
The CVE-20180708 Elmo version was evaluated against AlphaZero in the same conditions as the 2017 CSA Championship. This version had 44 CPU cores and 32GB of hash size, the same as Stockfish. AlphaZero won 91.2% of the games when played against Stockfish in sente, but the game was decided by one point. AlphaZero’s performance was impressive even before the game against Elmo began.
To train the AI for chess, the DeepMind team gave the machine some basic chess rules, and let it play against itself. The computer’s developers let it learn and experiment on its own, much like a newborn. AlphaZero was so experimental during the first test, that it didn’t follow any strategy. It tried every possible combination. It crawled slowly to checkmate and then completed another game. After nine hours of practice, AlphaZero could beat Stockfish, the world’s top chess engine.
After four hours of training, AlphaZero defeated StockFisher in a series of simulated games. The AI won in chess as well as shogi. After 30 hours of training, AlphaZero beat AlphaGo Lee-9 at Go. In all of these games, AlphaZero’s training algorithm consistently achieved similar levels of performance, which suggests that it is highly repeatable.
Evaluation of AlphaZero’s performance in chess
AlphaZero outperformed its human counterparts in a study that involved 57 games of chess and shogi. The subtree had eight million training steps. The results showed that AlphaZero could beat AlphaGo Zero nine in 61% of the games, and it recovered performance from an algorithm that exploited board symmetries and generated eight times as much data as AlphaGo Zero 9.
AlphaZero was significantly faster than Stockfish in all three tests and is almost three times faster than Stockfish. AlphaZero’s strength is primarily due to the fact that it uses a Monte Carlo tree search instead of the traditional heuristic method, examining 60,000 positions per second instead of the 60 million positions of Stockfish. AlphaZero also has four TPUs while Stockfish uses two.
AlphaZero uses binary planes and one hot encoding as input features. These planes represent counts, opponent, and policy. The program represents the game in the form of a stack, which corresponds to the entries in the table. To make the algorithm better, AlphaZero uses selected statistical data in both chess and shogi.
A key difference between AlphaZero and Go is its style. AlphaZero’s best moves occur in the first part of the game. The program does not calculate every outcome and will often sacrifice material early in the game. This decision will ultimately benefit the machine over the long-term. AlphaZero’s Go and chess Shogi matches will prove its superiority in both of these games.
Comparison with AlphaGo Lee
Lee Sedol is currently 3-1 against AlphaGo, the AI computer program. The match has been a landmark for artificial intelligence and its future work. It is a test of AI’s current capabilities and society’s reaction to artificial intelligence. The AI team at Google DeepMind stated that AlphaGo’s weaknesses were revealed in the match and that they would improve the program in future. But who will benefit from such a match?
In a paper published in Nature, the team revealed that AlphaGo ate up around 50,000 times more power than Lee Sedol. The difference is not significant enough to be considered comparable. It is important to note that the two computer systems were tested on very different machines. For the AlphaGo version, there were forty search threads and 280 graphics cards. Each move was given a time limit of five seconds.
Two teams played the games: an AI and a person. The AI beat Lee Sedol in the first match. However, the human player couldn’t win because he was too tired. Lee Sedol was playing against AlphaGo in his first game. Lee Sedol won the match in an informal format against AlphaGo in a second game. Despite the fact that Lee Sedol resigned, AlphaGo didn’t feel any pain or remorse. It was simply a result of the network’s search process.
Evaluation of AlphaZero’s performance in shogi
The evaluation of AlphaZero’s performance in both chess and shogi revealed a few striking features. Its MCTS search engine searches around 80 thousand positions per minute in chess and 40 000 positions per second for shogi. This is faster than Stockfish’s alpha-beta algorithm. Although this is significantly less than Stockfish’s 70 million position-per-second speed, the program compensates for this by using a deep neural network to focus on the most promising variations. These two features together make AlphaZero an arguably more human-like approach to search.
The researchers used head-to–head matches and a variety human opening positions to measure AlphaZero’s performance in chess. The time control used in the evaluations was a fraction that of 2017’s CSA World Championship. AlphaZero was better than Elmo in shogi with a time control of only one-fifth of the time.
A generalised reinforcement learning approach is used in the evaluation of AlphaZero’s performance in both chess and shogi. AlphaZero started with random play, and had no domain knowledge. It beat two world champion programs in less than 24 hours. It was called “the first superhuman program” in these difficult domains. The research team is working to refine and improve the program.