Chess game
A few months ago, a team of Google Deepmind researchers updated the article. Acidation Scheme with large -scale transformers: a case study on chessWhich did not pay any attention to anyone. It presents an artificial intelligence system (AI) chessBased on Transformer (neuronal network) Large scale, which plays the role of chess.
Without memory
To date, all chess programs were based on the complex search processes of the best movement in the tree that add the following plays and after reversal, often with a depth of more than 20 future movements. This, in association with the application of tasks, allows them to evaluate the benefits or disadvantages of each potential decision, to find the best game.
While the fundamental thing for all AIs was memory till now, Chessbench is the beginning of a new era, in which machines have been trained in planning works in which memory is useless.
Leela chess zero And Dry fish They are, today, the best chess engine, exponent of excellence in the game obtained with the first approach.
Both use learning based on neuronal networks from different ways: Leela chess uses reinforcement through zero auto games, and stockfish integrates a neuronal network for status assessment and decision -making improvement .
But Cheogle Deepmind Chessbench has just introduced a new model that can exceed all boundaries. As we have mentioned above, this is an AI system Transformer (Such as Burt and GPT) whose most advanced version is up to 270 million parameters.
Training
For their training and learning, 10 million chess games have been used LikesOpen source platforms to play online chess, in which thousands of human users worldwide connect daily. From these items, more than 15 billion anotations evaluating each situation and its best action (movement) are generated. It is evaluated Stockfish 16Whose search engine determines the best possible movement in a specific situation.
What makes chess
Training has allowed the system to learn to predict action values ​​for new positions with sufficient accuracy. This is a non -trivial normalization: it performs without the clear discoveries of the following plays and reversal.
What the result makes incredible is an AI system capable of playing chess at that level of quality.
Chess plays without the best sports process, only under the prediction of the next. This is the dream of any chess player: look at a position and recognize the best next movement by the configuration of the pieces.
It gives rise to a recurring question in debate about large -scale transformer and great model of language or LLM, such as behind a chat: Are they only great memormers or they can cause cause?
Chess
Chess requires a combination of strategy, strategy and prognosis. These are the main elements in plan, logic and decision -making processes.
Event Sourcing: To develop better?The model used by the Chesbench has 270 million parameters (2.7×10¹), while the number of potential chess plays is very high, about 10>. In this game, large -scale memoirs are useless. In addition to the complexity of the strategies involved, there is a vast variety of potential plays.
Chess may allow an accurate evaluation of the ability of the AI ​​model to be normalization and favorable for new conditions.
Chess has learned to recognize deep and complex patterns within the game of chess. This ability allows you to repeat your success even in situations not during your training.
The fact that you can identify and apply the winning strategies in not already viewed positions reveal the level of logic. It is beyond simple memoirs; This cannot only be held responsible for the coding of the next best play in its parameter space.
Chess Bet Records
Unlike other themes, chess is an objective method for estimating a player’s force: ELO score system. In narrowness the Blitz (Quick chess, with a game between 3 and 5 minutes per player for the whole game), only 15 people in the world arrived by chess over the lyl -lickbenth forum over 2 895 points. In the list Blitz Top 200 It can be seen that top 100 Keep Aloe Seema in 2 784.
Elo represents the level of a great teacher who arrived by Chessbnch. Solve challenging chess positions. In addition, it is playing against people in amazing Aloe Lich.
Remove alphabet
Chessbnch is still away from competition with AI AlphazeroThe level of which is above the world’s best human player. Perform in joint depth analysis with learning reinforcement, which allows you to learn to play against you.
Currently, chessbnch is a significantly good approach to the algorithm based on the search of stockfish. The correct distillation is still far from the scope of an AI system based on the observation of the situation that uses analysis of positions without search processes.
Generalization in learning
This makes the chasbench a sufficient initial point to investigate the future. For example, it will be interesting to know which performance will offer to play Fisher Modelity, that is, it is randomly located in the first line with large pieces. This increases the state of change of positions at the beginning of the departure, which requires great ability for generalization, as well as learning to learn the function of each piece and its legal movement.
recently In the initial position using the GPT4, an initial study has been done by exchanging horses for alphabets.And this reflects boundaries in these new situations for LLM. The study concludes that these scenarios have a lot to move forward and do more,
Imagination procedures
Judia pearlThe AI ​​researcher honored the Turing Award, the background scenarios, associates the situations that have not been seen in the universe currently seen by human research, For human process of imaginationThey are imagined worlds that could have been.
Large -scale transformers are not only language -oriented: different regions have an increasing number of applications. LLM is already being integrated with smart roboticsPlaying an important role.
Anyway, LLM architecture will be inadequate to receive one General It reaches the human cognitive level. But neither the status of those who say that they are only memoirs seem true and they are called “stochastic parrots”. This is a highly criticism.
The creative process is no longer unique to humansPlanning, robotics between other large -scale transformer skills, mathematical logic and continuous learning with the processing of sensory abilities of language processing, will affect everything known.
It will be interesting to know and understand how much the imagination of chess is.
The article has been done for 30 years in the ‘Machine Learning’ E -IA Field for 30 years in collaboration with a doctor, businessman and executive of Jose Louis Flores, Economics and Mathematicians.
Semgrep’s response: OpenGrep Open Code Analysis Ensures Tools
