Repository navigation
The first FunComp with a neural network evaluation (NNUE), trained only on FunComp's own games.
Strength. About 3640 on the CCRL Blitz scale: 3642 ± 8 from a single-threaded gauntlet of 2,000 games at 10+0.1 against 20 rated engines (CCRL Blitz 3492-3793), about 360 above 2.0 measured the same way (3282 ± 10). Against 2.0 directly: +484 ± 33 Elo in 1,000 self-play games from balanced openings (8+0.08: 888 wins, 108 draws, 4 losses). An estimate on our hardware, not an official CCRL rating.
Binaries.
funcomp-3.0-x86-64-v3.exe: for CPUs with AVX2 and BMI2 (Intel Haswell, AMD Zen and later). The faster build.
funcomp-3.0-x86-64.exe: for any 64-bit x86 CPU.
Both search identically (bench 1187868 nodes), are statically linked, and carry the network inside the executable: nothing to download or set.
New UCI option. EvalFile (default : the network built into the binary). A path loads another network; switches to the hand-written evaluation of 2.0.
What's new.
The network. 768 piece-square inputs for each of 7 king buckets (by the king's square, mirrored so the king stands on the queen's side), a 1,024-wide hidden layer for each side to move, squared clipped ReLU, and 8 output buckets by material. Updated incrementally move by move, with a cache of each king bucket's last accumulator, so a king move doesn't rebuild it from scratch.
Its data. 1.44 billion positions from FunComp's own self-play in three rounds, each labelled by FunComp's own search (round 1 with 2.0's hand-written evaluation, rounds 2 and 3 with the network of the round before), blended with the games' results. Openings: random plies from the standard position, Chess324 starts (Larry Kaufman's variant), and (round 3) positions from the Lichess open database (CC0), used as starting positions only. No other engine's games, evaluations or networks were used.
Its training. Our own trainer on the bullet library (GPU), 120 passes of 100 million positions. The width, king buckets, later layers and input features were chosen by games: matches at fixed nodes against a reference network, then timed matches, then a round robin of the three finalists at 2+1 (a tie; the simplest and fastest, without later layers, ships).
Search. Two features carried over from the 2.0-era tuning branch and kept after tests: correction history by the previous move (continuation correction), and hindsight depth (a node whose parent was reduced gets a ply back if the parent's move cost nothing, or one less if it cost a lot). The whole search re-tuned by SPSA for the network (2,000 iterations, 108 parameters).
Speed. The network's output in AVX2: about 2.7 million nodes a second on one thread of a Ryzen 9 5900X for the x86-64-v3 build (bench), about 1.5 million for the portable x86-64 build; 2.0's hand-written evaluation searched about 3.6 million.
Checked before release.
A stress run of 8,000 games on both builds (bullet, 1 s + 0.01 s, and 40 moves in 10 s, 24 at a time): no crashes, disconnects or illegal moves; no time losses at 40 moves in 10 s (see the limitations for bullet).
Threads with tablebases in bullet (4 threads, Syzygy, 2+0.02, five games at a time), 2,000 games: no failures.
No time losses in the 5,000 timed games of the gauntlet, the tune's confirmation and the match against 2.0.
Known limitations.
No Chess960.
In bullet with every hardware thread busy (1 s + 0.01 s, 24 games on 24 hardware threads), about 3 games in 1,000 were lost on time, by up to 61 ms, most likely to the operating system's scheduling: raise Move Overhead (default 30 ms) in such setups.
Next. Round 4's data (in generation: better labels from a deeper search and a second, backward pass over each game; more positions), then a network chosen by forecasts from the learning curves, and the learned guide for the search's pruning.