|Top Previous Next|
With Position Learning Houdini will remember advantageous principal variations it has analyzed by storing them in a learning disk file, for reuse in later games or analysis. The feature is completely transparent to the user, once Learning is enabled the data will be saved and loaded automatically.
The Position Learning is available in analysis and match play. But remember that it will only kick in for certain positions in which there is something to learn, i.e. when the score is significantly above the static evaluation of the position. In quiet, equal positions there's usually not much to learn.
Position Learning works best for positions where a good move exists that raises the score significantly above the statical evaluation. Learning is less useful and sometimes counterproductive for positions where many more or less equivalent moves exist and the goal is to avoid bad moves. To improve the efficiency of the learning the feature only kicks in when a certain positive scoring threshold is exceeded. By default this Learning Threshold is put at 10 centi-pawn (0.1 pawn).
Although the learning file grows very slowly, make sure not to allow the file to grow too large as the automatic loading may slow down the engine at the start of the analysis.
The impact of Position Learning can be illustrated as follows.
When running the Eigenmann Ending Test Suite with 20 seconds/move on a single-core laptop producing about 1,000 kN/s, Houdini finds 77 out of 100 solutions, the total solution time is 12:06. The test run produces a Learning File of about 27 kB.
Running the Test Suite a second time demonstrates the profit of the Learning File. Houdini now finds 80 solutions and the total solution time drops to 9:42.
When Learning is enabled the learning file will automatically be loaded at the start of the analysis, and further analysis will be appended to the file.
When Learning is enabled, automatically the Never Clear Hash option will be activated as clearing the hash would be counterproductive for the learning.
File name for saving the learning data.
A full file name is required, for example C:\Chess\Learn001.dat.
By default Houdini will use the Learn.dat file in the "My Documents" folder of the current user.
The learning file will be updated at the start of a new game or when you close the engine or GUI.
Min 0, Max 200, Default 10.
The Learning Threshold is the minimum positive score of a principal variation that is saved.
When Learning Threshold is larger the learning file will grow slower, as fewer moves will satisfy the condition..