3.7c

Version 3.7b is still weaker than version 3.0. These changes aim to restore the ELO after the new features and changes after 3.0:

• Keep Continuation History tables between the searches and not clear them, like in version 3.0.
• Enable again the LMR reduction's skip rate with a high value in order to not affect CPU performance.

3.7b

Less aggressive randomization of move order before sorting.

3.7a

• Change the condition for randomizing the non-attacking moves. Now they are always randomized.

3.7

• SMP version:
• 1. On https://tcec-chess.com/ S24, there is still NPS scaling issue, because the achieved value is 5-10M nps and the expected is 30M nps.
• 2. The issue is reproducible on CentOS Stream release 8 and Java OpenJDK 1.8.0.
• 3. The tests show that with up to 64 cores (128 logical processors) it still achieves more stable and deeper depth, for a fixed time, with only 1 Transposition Table, although the NPS is between 3 and 6 times smaller.
• 4. A new UCI option is introduced for SMP version - CountTranspositionTables.
• 5. On CentOS Stream release 8 and Java OpenJDK 1.8.0, with 8 Transposition Tables the SMP search achieves similar depth and time, so there is sense to set it not to be 1, only when there is a NPS scaling issue. In such cases it could be set up to SQRT(THREADS_COUNT) Transposition Tables.
• 6. More details about the tests of the SMP version, could be found here: https://github.com/bagaturchess/Bagatur/tree/master/Sources/Search#readme
• 7. Bug fixing and optimizations inside the bagaturchess.search.impl.rootsearch.parallel.SearchersInfo class.
• 8. SMP version: Code clean up and simplification
• Single Core version
• 1. Randomize the non-attacking moves, before moves' ordering and sorting, in order to benefit from the search's undeterminism and minimize error.

3.6b

• TBs: DTZ or DTM check improved. Still not 100% accurate - alone king condition used. Still direct DTM from native syzygy lib is not used, because of JNI/C types related issues, which need more time.
• TBs: Syzygy native lib changes and JNI integration code adoption - methods types are incorrect. How Value has to be converted to int?
• Search revision: move ordering based on the LMR beta cutoff move's ration. Added clear method in History Tables. CLEAR_TABLES_ON_NEW_SEARCH flag introduced and set to false.
• Search revision: enable LMP and LMR for all nodes (without statistical constraints - in this case "faster" is over "too accurate")

3.6a

• Bugfixing: Syzygy TB support - DTZ handling
• Optimization: Decision whether to use or not LMR based on below alpha statistics of previous LMR searches

3.6

• Add Syzygy TB support
• Bugfixing: Bagatur evaluation's jumps in SMP version

3.5

Code improvements and bugfixing in regards to UCI options handling

3.4h

• Bugfixing: When setting UCI options, which need re-creation of search objects, use their direct names, without converting to lower case.
• Optimization: create search objects only once, after all UCI options are set.

3.4g

• Bugfixing (not related with FRC), of an issue introduced with version 3.4f. It is related with the extraction of moves from Transposition Table and building the PV to show as output.
• Bugfixing (not related with FRC), sometimes in mate positions, after repetition of move 2 times, Bagatur evaluation jumps between 0 and +-mate score.