Caissatrack

Track the evaluations of the 100k most popular positions in Caissabase on chessdb.cn (cdb). The positions can be found in caissa_sorted_100000.epd, where they appear in sorted order according to popularity. For example, the starting position appears first. The average depth of these positions is 14.6 plies, with the deepest one 42 plies away from the starting position. The positions have 30.3 pieces on average, and none has fewer than 18 pieces on the board.
Thanks to Joost VandeVondele for preparing caissa_sorted_100000.epd in July 2023.

The file caissa_sorted_100000_cdbpv.epd contains the current cdb evaluations and PVs for each position. It is created daily with the help of the script cdbbulkpv.py from cdblib, and the obtained statistics are written to caissatrack.csv. Moreover, each day the thousand positions with the currently shortest PVs on cdb (ignoring PVs ending in a two-fold repetition) are written to caissa_daily_shortest.epd, and the positions with absolute evaluations in the interval [80, 110] are written to caissa_daily_edgy.epd.

---

---

---

Progress

The following graphs attempt to measure the progress cdb makes in exploring and evaluating the positions in caissa_sorted_100000.epd.
They plot the evolutions in time of the two (daily) indicators

$$E = \frac{1}{N} \sum_i \left(\min\{\frac{|e_i|}{100},2\} - 1\right)^2 \qquad \text{and} \qquad D=\sum_i \frac{1}{d_i},$$

where $(e_i, d_i)$ are the evaluation and depth values for the 100K positions, with the convention that $d_i = \infty$ if the position's PV ends in a terminal leaf (2-fold repetition, 50 moves rule, stalemate, checkmate or 7men EGTB). $E$ measures how certain cdb's evaluations are, while $D$ simply sums the inverses of the lengths of the non-resolved PVs. Note that as cdb (slowly) approaches the 32men EGTB, $E$ should converge to 1, while $D$ should converge to 0.
In addition the graphs also show the evolution of the total number of "on edge" positions. The first graph shows all available data points, while the second graph depicts only the 100 most recent data points.

---

Get involved

If you want to help improve the coverage of these positions on cdb, you could manually or systematically explore them on chessdb.cn. As an example for the latter, you could clone this and Joost VandeVondele's repo cdbexplore via

git clone https://github.com/robertnurnberg/caissatrack && git clone https://github.com/vondele/cdbexplore && pip install -r cdbexplore/requirements.txt

and then from within the cloned caissatrack repo either run

git pull && cat caissa_daily_shortest.epd caissa_daily_edgy.epd > caissa_daily.epd && python ../cdbexplore/cdbbulksearch.py --bulkConcurrency 16 --forever --shuffle caissa_daily.epd

occasionally, or run

python ../cdbexplore/cdbbulksearch.py --bulkConcurrency 16 --forever --depthLimit 10 --shuffle caissa_sorted_100000.epd

as a long-term job.

An automated solution is to run the script launch_caissa_daily.sh automatically via .crontab entries of the form

@reboot sleep 20 && /path_to_script/launch_caissa_daily.sh
55 6 * * * cd git/caissatrack && git pull

where the second entry is really only needed for server-like machines that run 24/7.

---

Get in touch

To discuss anything cdb related, and to help cdb grow at a healthy pace, join other (computer) chess enthusiasts at the chessdbcn channel on the Stockfish discord server.