Chess programs can differ in depth of search or in the evaluation function applied to leaf nodes or both. Over the past 10 years, the notion that the principal way to strengthen a chess program is to improve its depth of search has held sway. Improving depth of search undoubtedly does improve a program's strength. However, projections of potential gain have time and again been found to overestimate the actual gain. We examine the notion that it is possible to project the playing strength of chess programs by having different versions of the same program (differing only in depth of search) play each other.
We're not being replaced by AI. My chess loss in 1997 to IBM supercomputer Deep Blue was a victory for its human creators and mankind, not triumph of machine over man. In the same way, machine-generated insight adds to ours, extending our intelligence the way a telescope extends our vision. We aren't close to creating machines that think for themselves, with the awareness and self-determination that implies. Our machines are still entirely dependent on us to define every aspect of their capabilities and purpose, even as they master increasingly sophisticated tasks.
What is being learned by superhuman neural network agents such as AlphaZero? This question is of both scientific and practical interest. If the representations of strong neural networks bear no resemblance to human concepts, our ability to understand faithful explanations of their decisions will be restricted, ultimately limiting what we can achieve with neural network interpretability. In this work we provide evidence that human knowledge is acquired by the AlphaZero neural network as it trains on the game of chess. By probing for a broad range of human chess concepts we show when and where these concepts are represented in the AlphaZero network.
We present an end-to-end learning method for chess, relying on deep neural networks. Without any a priori knowledge, in particular without any knowledge regarding the rules of chess, a deep neural network is trained using a combination of unsupervised pretraining and supervised training. The unsupervised training extracts high level features from a given position, and the supervised training learns to compare two chess positions and select the more favorable one. The training relies entirely on datasets of several million chess games, and no further domain specific knowledge is incorporated. The experiments show that the resulting neural network (referred to as DeepChess) is on a par with state-of-the-art chess playing programs, which have been developed through many years of manual feature selection and tuning. DeepChess is the first end-to-end machine learning-based method that results in a grandmaster-level chess playing performance.
The basic paradigm that computer programs employ is known as "search and evaluate." Their static evaluation is arguably more primitive than the perceptual one of humans. Yet the intelligence emerging from them is phenomenal. A human spectator is not able to tell the difference between a brilliant computer game and one played by Kasparov. Chess played by today's machines looks extraordinary, full of imagination and creativity.