Li Shishi 78 L11 fourth hand dug by everyone as the "hand of God", in the policy of the network output in DarkForest ranked 31, and ranked 10 by J11. So I think there might be AlphaGo not count to this step. If the person did not count under the hand machine to chess, Monte Carlo (MCTS) will clear the search tree, and then re-start the search, we should not draw conclusions too quickly. Li Zhe sixth tell me K10 seconds under the hand, it might be time management subsystem procedural flaws at the time of the search tree emptied, thus prematurely search results returned. MCTS at the beginning of the search, because the number of simulations are not enough, each step of winning the variance is very large, so the return is not good enough with a method such as K10 is normal (in this DarkForest inside the top four). This relatively easy to fix.

Another possibility is that, AlphaGo valuation network problem. Because the right to re-valuation of the network is 0.5, regardless of the child quickly go from a situation repeat how many times its weight is 0.5. For a situation, the valuation of the network only get a number, and go down from this situation the child will get a lot more than the number to go after, statistically should be more important, but AlphaGo not think so, both sides claim to be winning after each direct in half the average. So if the results of a valuation of the network situation was not right, it will greatly affect the situation of the winning estimate. We get a lot of attention here because according to article number, leaf nodes have accumulated after a certain number of plate (40) to expand, rather than the first visit had begun to improve DCNN efficiency. DarkForest not used valuation network, after digging correctly L11 L12 and L10 returned both hands should be, according to Li Zhe sixth that these are the right hand should be, which indirectly supports this inference. AlphaGo after 87 hands only realize that they have lagged far behind, it may also be due to the same problems, such as the big black dragon on the right to see survival.

Why does the valuation of the network will be problems? It may be used for self-learning and training network valuation (Self-Play) sample distribution blind spots. In order to increase the production rate of the sample, the sample AlphaGo is self-learning by two purely DCNN each stroke to generate (no search), and DCNN at chess because it is a pure pattern recognition, a big problem is that life and death is not correct, often erupted in a stupid move inside. If black and white two parties have committed, regardless of life and death problems, for example, then the white one lucky win, and that the valuation of the network will think just stupid move white situation is good. Such valuation network will be infected with the same problems, misjudgments in the disk of the complex situation in the killing. If this situation is hard to deal with, AlphaGo next Board may also have the same problem. Here you can see, the computer itself is not to rely on exhaustive chess, Go, after all, is too complicated, prune every step, from the current situation carefully near Shear (with DCNN), the current situation is far from fast cutting (fast go child) until the final outcome so far obtained. Pruning is directly related to the level of chess, DCNN just a bigger picture of a very good means of pruning, it will blind spot by losing move reflected.

About DCNN + MCTS robbery. MCTS is a global first, because the estimated fraction of ko itself and other situations in the program appears to be no essential difference, just a move it. Robbery particularity in DarkForest performance hit may have robbed mentioning happens, DCNN often very high probability (0.8 or more) returns ko capture this hand. Probable cause is that robbery is a point as a separate feature input, so DCNN to learn it and output (ko) strong correlation. So MCTS search would strongly favor this hand. In many cases this is true, but sometimes it is very small and can be disregarded robbery, robbery or across two or more of a need to give up, that the "robbery will lift," the bias will be to search for trouble. Sometimes the computer serial robbery repeatedly mention this reason. AlphaGo may have this problem, or is the reverse problem (for example, ko probability is very small), so that everyone will feel it when playing chess in the open to avoid robbery or miscalculation changes contained in the robbery.

About Horizon effect (Horizon Effect). As the country's AI which have this effect, for example, to search only 10 steps to calculate the others after their end after eating, and then use a simple additive method to estimate the disk find themselves under particular plurality cool, I feel that this branch particularly good. In fact, their progress downward step after others have been eaten, or fall into the trap, so mistakenly thinking face value. But go because each simulation are both in the end, probably before using 30 steps DCNN, after the child is to use the fast walking, although there are gaps in quality child go, but the wrong direction, so to some extent the horizon effect on is weakened. And this AlphaGo mistakes within 20 steps, which should still DCNN range, so the possibility is relatively low horizon effect.