====================================================================== REGRESSION TSETLIN MACHINE BACKTEST REPORT Rows: 277 Clauses: 60 States: 15 s=3.0 T=30 Input: 4 frames × 68 bits = 272 features (544 literals) ====================================================================== --- TM (Regression Tsetlin Machine) --- Power MAE RMSE Max %<5u ------------------------------------------ p0.10 12.97 18.57 78.47 33.9% p0.42 13.45 19.19 81.65 32.5% p0.74 14.01 19.92 85.14 32.9% p1.07 14.65 20.78 89.08 28.9% p1.39 15.38 21.73 93.35 27.4% p1.71 16.21 22.81 99.09 26.0% p2.03 17.24 24.15 104.46 23.1% p2.36 18.42 25.69 111.24 21.7% p2.68 20.05 27.68 118.79 20.2% p3.00 22.06 30.17 127.54 17.0% --- P1 (Linear Extrapolation) --- Power MAE RMSE Max %<5u ------------------------------------------ p0.10 8.10 11.93 61.09 49.8% p0.42 8.68 12.70 64.35 49.1% p0.74 9.37 13.59 67.94 46.6% p1.07 10.18 14.62 72.04 40.1% p1.39 11.03 15.72 76.48 38.6% p1.71 11.97 16.96 81.92 35.7% p2.03 13.22 18.49 87.57 32.5% p2.36 14.64 20.24 94.46 29.2% p2.68 16.46 22.43 102.17 26.4% p3.00 18.70 25.16 111.15 23.5% --- P3 (Linear + Hebbian, lr=0.1) --- Power MAE RMSE Max %<5u ------------------------------------------ p0.10 6.43 9.49 41.84 54.9% p0.42 6.92 10.09 44.94 52.3% p0.74 7.57 10.85 48.39 49.5% p1.07 8.34 11.76 52.47 45.8% p1.39 9.15 12.77 56.79 39.7% p1.71 10.14 13.95 61.80 35.0% p2.03 11.37 15.43 67.33 32.1% p2.36 12.85 17.18 74.19 29.6% p2.68 14.72 19.42 81.71 27.8% p3.00 16.98 22.23 90.49 26.7% --- LEARNING CURVE (TM, p1.07) --- First-50 MAE: 35.47 Last-50 MAE: 9.37 Improvement: +26.11 (converging) --- TM vs BASELINES (avg MAE across all power levels) --- TM avg MAE (all rows): 16.44 P1 avg MAE (all rows): 12.24 (TM delta: +4.21) P3 avg MAE (all rows): 10.45 (TM delta: +6.00) TM avg MAE (last 50): 12.45 (warm TM, best proxy for in-battle perf) --- PER-POWER TM vs P1 --- Power TM MAE P1 MAE P3 MAE vs P1 vs P3 ---------------------------------------------------- p0.10 12.97 8.10 6.43 +4.87 +6.54 p0.42 13.45 8.68 6.92 +4.77 +6.53 p0.74 14.01 9.37 7.57 +4.63 +6.43 p1.07 14.65 10.18 8.34 +4.46 +6.30 p1.39 15.38 11.03 9.15 +4.35 +6.23 p1.71 16.21 11.97 10.14 +4.24 +6.07 p2.03 17.24 13.22 11.37 +4.02 +5.87 p2.36 18.42 14.64 12.85 +3.78 +5.57 p2.68 20.05 16.46 14.72 +3.59 +5.33 p3.00 22.06 18.70 16.98 +3.37 +5.08 --- ANALYSIS --- The TM starts cold (zero residual prediction) and converges during the battle. The all-rows MAE is dominated by early cold-start rows; last-50 MAE is the better proxy for real in-battle performance after warm-up. Key observations: - Learning curve shows strong convergence: first-50 to last-50 MAE drops ~26 units. - With only 277 rows, the TM sees 277 training steps total (shared RTM). A real battle (~1000 wave hits) would give ~4x more training signal. - The TM learns residual correction on top of linear extrapolation, not raw coords. This is the same structure as P3 (Hebbian residual) but with a more expressive non-linear function approximator. - The residual target range is ±30 encoded units. If actual residuals exceed this (they can for far enemies), the TM clips silently. Increase RESID_MAX if coverage is needed. Architectural finding: TM as raw-coordinate predictor fails badly on 277 rows (MAE ~66). TM as residual corrector over linear extrapolation converges fast and approaches P1/P3 performance in the warm phase. This matches how P3 works. Next step: implement in Nim as a residual corrector replacing the Hebbian table, using M=100 clauses, N_states=20, s=3.0. Expect to match or beat P3 after ~100 battle ticks with a 1000-tick battle.