Komodo 11.2.2

Table Memory

This parameter determines how much memory, in megabytes, Komodo will allocate to various internal tables such as the pawn structure hash table, evaluation cache and others. These are used by Komodo’s evaluation to prevent doing duplicate work. If you allocate too much or too little memory it can negatively impact your performance. The default value of 128 is probably close to optimal for fast play on most machines cira 2013 but you may want to experiment with raising this value if you have a high performance machine. For slower games (or analysis) and/or for machines with many cores, higher values are appropriate. We have raised the upper limit in Komodo 9.2 and higher to 1024 megabytes since this may be appropriate for TCEC or similar use with many cores and long time limits. We used 512 in our recent tournament victories.

Null Move Pruning

Like LMR, this is also a technique to aggressively prune moves from the search tree. Komodo plays much stronger with Null Move Pruning turned ON, but occasionally things can be missed, especially very deep threats and zugzwang positions. This option allows you to experiment in the case that you have a difficult position which cannot be resolved and want to see if a much shallower, but more thorough, search would help.

Null Scale

Null Scale lets you modify how aggressively Komodo reduces the depth when trying Null Move searches. Higher values will reduce more, letting Komodo search deeper at the risk of missing some things. Lower values will make Komodo play more safely with less reduction in Null Move searches. Null Move Pruning must be on for Null Scale to have any effect. A Null Scale of 28 has tested best and is the default.

Overhead (milliseconds)

This is designed to compensate for slow user interfaces or slow internet connections, where a fraction of a second can be lost in the processing and transferring of information back and forth from the engine to the interface or interface to a chess server. The value is in milliseconds (1/1000th of a second). This should probably be left alone unless you see the program starting to forfeit games due to exhausting its time. If you see such time forfeits, increase this value. The default is 30. For internet play where lag is an issue a value of 60 is a good first try, but may have to go to 100 or even higher if you see time losses. For games with fairly long time control, 50 is a good value for general use just to be safe since the lost time will not be noticeable at these levels.

Contempt

A positive Contempt is used by the computer’s evaluation to discourage (or encourage if negative) piece trades, blocked pawn structures, and repetition draws early in the game. The default value is 10 (in version 9.3 and higher) which should help against weaker opponents. Against near-equal opponents, set Contempt to 0 for best results. Against a human opponent, Contempt should be at least 25 even if he is the World Champion. Suggested values include 50 against grandmasters, 75 against masters, and 100 against strong amateurs. A good rule of thumb is to subtract the opponent's Elo rating from 3450 (estimated rating of Komodo 11 on a quadcore machine) and divide by 15. Higher Contempt values are especially appropriate when Komodo is giving a handicap. For handicap play, a good rule of thumb is to set Contempt for one third of the value of the initial position based on a one minute search in infinite analysis mode. For example, if giving knight odds, if an initial one minute analysis search shows Komodo as down by 3.3 pawns, set Contempt to 110, which is 110 centipawns, or a third of the disadvantage. Contempt works with the next option “White Contempt”. In analysis mode, Komodo internally sets Contempt to zero unless “White Contempt” is turned on.

Time Usage

This provides the ability to manipulate Komodo's time control heuristics, to make it allocate time more quickly or less quickly. The default is 0 which is what we believe is the ideal setting but remains to be tested at longer time controls. You can make it play faster (use less time) by using a negative number or you can make it allocate time more aggressively by setting it to a positive value. The range of values is between -9 and +9. It's very unlikely extreme values will play well, so we suggest experimenting with relatively modest changes to the default. The ideal value may depend on the time limit.

Syzygy Probe Depth

This determines the depth at which tablebases will be probed during search (note that the root position will always be probed. If 'Use Syzygy' is enabled and the number of pieces on the board is less than or equal to the 'Syzygy Probe Limit' [see below] or the number of pieces supported by your installation of the Syzygy bases, whichever is less). A higher number means that Komodo will wait longer to begin probing the tablebases for any particular search. The default depth is 2 half-plies.

Syzygy Probe Limit

This determines how many pieces need to be on the board before Komodo begins probing (even at the root). By default, this is set to 6 (the current upper limit of the Syzygy tablebases). Of course, if you have only the 3-4-5-man set, Komodo will only begin probing the bases when a 5-man position arises. Note 6 piece use can take up both a lot of drive storage and RAM during the search.

Smart Syzygy

When this is checked/on, Komodo will always check Syzygy TBs for (Syzygy Probe Limit - 1) pieces and under at all depths in the main search. Otherwise, it will only consult Syzygy when the Probe Depth and Probe Limit conditions are met. This feature is designed to allow Syzygy access from slower hard drives.

For instance, set Syzygy Probe Limit to 6 (the default if you have the 6-man tablebase files), Probe Depth to 10 and enable Smart Syzygy. At search depth 10 and higher, Komodo will use Syzygy for 6-piece positions. At lower search depths, it will only consult Syzygy when there are 5 pieces or fewer on the board.

Selectivity

You can adjust how aggressively Komodo prunes using the "Selectivity" parameter. Higher values will encourage Komodo to prune more at low search depths, leading to a deeper overall search at an increased risk of missing some moves. The default in Komodo 11.2.2 is 120. A setting of 250 will set selectivity to the maximum.

Reduction

This controls how deeply Komodo searches. Higher "Reduction" values will cause Komodo to reduce parts of the tree more, increasing overall search depth but possibly missing some good lines. The default is 0, meaning no adjustment to what we consider best. We are fairly confident that the best value on a single thread is close to zero, but we have not yet determined whether other values may be better on many threads, so feel free to experiment.

King Safety

This controls how highly Komodo scores king safety. The higher the value, the greater the effects of attacks on the king in Komodo's evaluation. The default is 73 in Komodo 11.2.2.

Persistent Hash

This is useful for long analysis. It allows you to save the current Hash Table to your hard drive, then reload it later. First, set a name for the file in the "Hash File Name" option. Most GUIs will require you to close the dialog box with an OK to “set” the name. You should then stop the search/analysis, return to the UCI parameters then, then click the "Save Hash to File" button. To later reload the saved information, first make sure the "Hash File Name" is for the file you want, then click "Load Hash from File". Once it loads, you may continue analysis. You must make sure the current hash filename is the same as what was used when you saved the file. Note that hash files can be very large and can take a lot of time to save and load, so make sure you have enough disk space to use them.

Minimal Reporting

This feature allows you to reduce the data Komodo outputs for the first X ply of a search. The default value of 0 does not suppress any program output. Value of 1 or more will suppress outputting the current move Komodo is searching. Higher numbers will suppress more output. For example, with a value of 8 Komodo will not print move sequences or scores for the first 8 ply. This can save on communication overhead with the GUI and is useful when Komodo is running over a network connection or is playing at fast time controls. The default value is 0, meaning all data is reported and nothing is suppressed.

Dynamism

Introduced in Komodo 9.4, Dynamism lets you control how important Komodo thinks the dynamic part of the evaluation is. The default is 120. Values under 120 reduce the dynamic evaluation, while values above 120 increase the dynamic evaluation. Lower values will make Komodo play more conservatively, while with higher values it will play riskier moves. More static parts of the evaluation like material and pawn structure are not changed by Dynamism. Although the parameter can be set from 0 to 400, single-digit values will cause bizarre play. A value around 80 will in general produce more realistic evaluations than the default 130, although it will weaken play by 20 to 30 Elo points. The maximum Dynamism value was raised at the request of a Komodo user for version 11.01 and newer.

Progress Threshold

Progress Threshold allows you to set the number of half moves played with no capture or pawn advance (per the 50 move rule) where Komodo starts pulling the evaluation score towards zero. We used 85 until mid-2016, then reduced it to 30 for later versions, which probably cost one or two elo but made Komodo a better analysis tool and shortens long clearly drawn endgames. The lower this value the sooner Komodo will start to see that no progress is being made and encourage it to take another line.

Variety

Variety lets you have Komodo play different lines from its default. A value of zero turns off Variety. With the setting at zero, Komodo will always play the best move it can find. A number from 1-20 allows Komodo to find different moves in some positions. The higher the number, the greater the variety of moves will be. Higher values should lower elo, although we could not detect any elo loss (tiny gain actually) after many thousands of ultra-fast games with a setting of 5 or 10, and less than one elo loss with the maximum setting of 20, with a three elo error margin in each case. Perhaps it was a bit lucky; but it seems that the weakening is not more than one elo per five points set, probably less. Note that when still in Opening Book (described below), the book determines what moves will be played.

Note that running Komodo on more than one thread automatically produces variety, even at the same depth, but the Variety option works even on one thread, and should increase variety on multi-threads.