Eubos v3.6

• Increase further the en prise piece malus component in evaluation
• Refactors to speed up performing and unperforming moves
• Refactor to remove disabled features that didn't increase playing strength and had an impediment on code/algorithm complexity.

Eubos v3.5

Release for 108th Amateur Tournament.

I disabled some features that I found to be a net detriment and also tuned the evaluation a little.

Eubos v3.4

Tweaks to time management strategy. Now uses more time earlier in the game. Also responds immediately when a forced move must be played. Minor refactors to some of the timing and validation code.

Eubos v3.3

Bug fix for history heuristic at long time controls, to prevent saturation of move history scores.

Eubos v3.2

Several minor tweaks and refactored the code for running EPD test suites.

Not expected to be significantly stronger than v3.1.

Eubos v3.1

Adds history heuristic to move ordering
Enable knight check threat to evaluation
Tune LMR move and futility thresholds
Disable passed pawn imbalance factor in evaluation
Package required JARs rather than extracting them into the Eubos.jar file

Eubos v3.0

This is a major release because I changed the primary search algorithm to NegaScout from NegaMax.
I also made an ammendment to the hash allocation, so that Eubos now uses the full hash allowance, rather than using the next lowest power of 2 to the maximum hash table size, as it did previously.
These two changes have improved the strength of Eubos, particularly at longer time controls.

Eubos v2.27

This release contains a major bug fix. A bug existed since Eubos 2.24 and perhaps earlier. The defect is more prevalent at longer time controls, as it related to the hash table, which is more important in longer searches. The best move that was returned could be legal, but spurious. It could result in not moving away an attacked piece, or even making a losing capture for no positional/mating benefit.

I have also, as a consequence of trying to corner the defect, added a validation search and extra diagnostic logging, which is switched off in the released binary.

Also as a consequence, I have revisited the search functions and made various refactors and improvements to aid maintenance of the code. The best of these being reworking the move iterator for the lazy move generation to be more logical and efficient.

I have also added storage of extended/quiescent search transposition data in the hash table.

Some more gory details on the defect...
The processing that returned the best move failed to check the depth of the transposition table entry, which could mean that if a hash entry was overwritten during the search, after backing up a hash move (for example, if the root hash got overwritten by the replacement scheme and then a transposition of moves meant the same position was reached further down the search tree, but searched to a lower depth) then the best move could be legal but spurious. The regular search correctly checked the depth.

Eubos v2.26

Bug fix for CCRL 104th Amateur Series Division 10

This release works around an issue whereby the move returned in the UCI best move command could be different than the computed principal variation. That is always bad, sometimes awfully bad. The defect seems to be most prevalent at longer time controls, it is possible its impact is fairly neglible in bullet and faster blitz time controls.

Eubos v2.25

Primarily, reworked the use of the hash table in quiescent search. This has yielded a significant Elo improvement. Eubos is now stronger than Dragon 4.6. Latest result on release binary was a 50.7% win rate over 22,500 games.