Projects

Rodent-V

GoChessNNUEAssemblySystems Programming

Rodent-V is a 3500+ ELO superhuman UCI chess engine written in Go by Naman Thanki and Pawel Koziol, based on Sungorus 1.4 by Pablo Vazquez. Recognized as the world’s strongest Golang chess engine and ranked in the top 100 worldwide, it pairs master-level strength with a deeply customizable personality system.

Strength & Neural Networks

  • Default Net (3500 CELO): A bullet-trained 768 -> (512*2) -> 1 architecture with horizontal mirroring, evaluated via custom x86-64 AVX2 assembly kernels (nnue_avx2_amd64.s) for high-throughput SIMD inference in Go.
  • Tal Net (3300 CELO): Ranked as the world’s 9th most aggressive engine, achieving an EAS (Engine Aggression Score) that ranks alongside Patricia, specializing in relentless sacrifices, king-hunts, and sharp tactical imbalances.
  • Hybrid Evaluation: Capable of running pure NNUE, pure handcrafted evaluation (HCE), or a weighted blend of both.
  • Handcrafted Evaluation (HCE): Covers material, piece-square tables (PST), mobility, king safety, passed pawns, pawn structure, drawish endgames, and interpolated game phases, with dynamic PST adjustments driven by central pawn structures.
  • Training & Tuning: Features a built-in tuner for HCE parameters and an automated datagen pipeline for NNUE training.
  • Personalities: Fully configurable playing tendencies and evaluation weights driven by external personality files.

Search & Selectivity

  • Search Framework: Iterative deepening with Principal Variation Search (PVS) and make/unmake board state updates.
  • Quiescence Search: Static Exchange Evaluation (SEE) filtering and check evasion generation.
  • Transposition Table: 4-bucket cluster hash table with entry aging.
  • Pruning & Reductions: Reverse futility pruning, razoring, dynamic null move pruning, late move reductions (LMR), late move pruning (LMP), and futility pruning.
  • Extensions: Stacked singular extensions and check extensions.
  • Move Ordering: Transposition table move, MVV-LVA / SEE captures, killer heuristics, history heuristic, 2-level continuation history, and correction history.

Source code: github.com/nescitus/Rodent-V