On Sat, Feb 12, 2011 at 1:46 AM, Christian Mascher <span dir="ltr"><<a href="mailto:christian.mascher@gmx.de">christian.mascher@gmx.de</a>></span> wrote:<br><div class="gmail_quote"><blockquote class="gmail_quote" style="margin: 0pt 0pt 0pt 0.8ex; border-left: 1px solid rgb(204, 204, 204); padding-left: 1ex;">
<br>
<br>
Interestingly the breakthrough achievement leading to todays much stronger programmes (only a couple of years back) did not come by mimicking human reasoning about the next move. Instead, the merit of a move is evaluated by a monte-carlo approach: The board is filled with random moves until the machine can tell who wins. This is repeated thousands of times and so by sheer statistics the relative value of the next move is assessed. This kind of evaluation process is absolutely not what a human player is doing or capable of doing in his head. So there is no hope of computers teaching us how to play better Go ;-), at least not in the sense of explaining to us why they made a certain move.<br>
<br></blockquote><div><br>My understanding is that the strongest Go programs use the Monte Carlo method, evaluating the various moves against vast databases of past professional games. So they are not really mimicking human reasoning, but are "data mining" the wisdom of humans. The programs can report back some info about why they made a certain move, usually in the form of "this move was made in 60% of pro games in a similar position that eventually led to a win".<br>
<br>In this sense, Go is proving to be a good example of the "new approach" to AI. For example, translation between foreign languages has undergone a similar revolution. 20 years ago, it was all about building expert systems with thousands of carefully crafted rules that capture the domain knowledge. These days, we just throw a large body of examples of human-translated documents at a data mining program, and let it work out new translations using statistical methods.<br>
</div></div>