SirStone
1ed7797cb6
- New guns: guess-factor (GF histogram), pattern-matcher (movement tape replay) - New modules: minimum-risk melee movement, spinning melee radar - New test bots: PatternMover, RandomMover, WaveSurfer - Fixed: FeedbackEvent now carries actualX/actualY for proper GF learning - Fixed: TM gun warmup gating + directional residuals - Fixed: circular gun integrated formula + multi-bin omega cache - Fixed: oscillator wall-bounce lockout - Fixed: phantom meteor perpendicular body orientation - 6/6 battle wins across all enemy types
41 lines
2.4 KiB
Plaintext
41 lines
2.4 KiB
Plaintext
================================================================================
|
||
ALTERNATIVE LEARNING METHODS BACKTEST
|
||
Dataset: 277 rows | Rep power: p1.07 | Hit threshold: 18.0px
|
||
Baselines: linear MAE≈12.24 | WiSARD K=12 MAE≈9.93
|
||
================================================================================
|
||
|
||
Method MAE Hit% F50 MAE L50 MAE vs Lin
|
||
--------------------------------------------------------------------------------
|
||
Linear (baseline) 10.18 80.1% 20.74 8.19 ---
|
||
WiSARD K=12 (reference) 9.93 ~55% --- --- -2.31
|
||
1. Echo State Net (res=512) 8.29 85.2% 15.76 2.11 -1.89
|
||
2. Kanerva SDM (addr=2000) 10.16 80.1% 20.74 8.19 -0.02
|
||
3. N-gram Markov (4×69 chunks) 10.11 80.5% 20.36 8.19 -0.07
|
||
4. Bloom Filter (8192 slots) 17.08 59.2% 23.67 13.15 +6.90
|
||
5. HDC (n_hd=2000, 32 cls) 65.33 0.0% 56.11 65.33 +55.15
|
||
6. RandSubspace (30×50bits) 8.31 90.6% 15.45 6.00 -1.88
|
||
7. WiSARD+Elig (k=12,tr=5) 8.24 90.3% 15.21 5.74 -1.94
|
||
|
||
================================================================================
|
||
RANKING by MAE (lower is better, rep power p1.07)
|
||
================================================================================
|
||
|
||
Rank Method MAE Hit% F50 L50
|
||
----------------------------------------------------------------------
|
||
1 7. WiSARD+Elig (k=12,tr=5) 8.24 90.3% 15.21 5.74
|
||
2 1. Echo State Net (res=512) 8.29 85.2% 15.76 2.11
|
||
3 6. RandSubspace (30×50bits) 8.31 90.6% 15.45 6.00
|
||
4 3. N-gram Markov (4×69 chunks) 10.11 80.5% 20.36 8.19
|
||
5 2. Kanerva SDM (addr=2000) 10.16 80.1% 20.74 8.19
|
||
6 4. Bloom Filter (8192 slots) 17.08 59.2% 23.67 13.15
|
||
7 5. HDC (n_hd=2000, 32 cls) 65.33 0.0% 56.11 65.33
|
||
|
||
Linear baseline: MAE=10.18 Hit=80.1%
|
||
WiSARD K=12 ref: MAE=9.93 Hit=~55%
|
||
|
||
Notes:
|
||
F50/L50 = MAE on first/last 50 samples (learning speed proxy).
|
||
Hit% = fraction within 18px (Robocode bullet half-width).
|
||
All methods: online, binary input, no gradients, no supervised labels.
|
||
Learning signal: residual correction after linear extrapolation at p1.07.
|