Task 3: counterfactual twin delib-v2 player (~15.14 +- 0.13)

braincraft/env3_player_ensemble_counterfactual_twin_delib_v2.py

@@ -0,0 +1,294 @@
+
+
+from __future__ import annotations
+
+from typing import Iterable, Tuple
+
+import numpy as np
+
+
+def relu(x):
+    return np.clip(x, a_min=0.0, a_max=None)
+
+
+def identity(x):
+    return x
+
+
+def ensemble_counterfactual_twin_delib_v2_player() -> Iterable[Tuple]:
+    n = 1000
+    p = 64
+    warmup = 1
+    leak = 1.0
+
+    hit_idx = 2 * p + 0
+    energy_idx = 2 * p + 1
+    bias_idx = 2 * p + 2
+
+    side_right_idx = int(np.clip(52, 0, p - 1))
+    side_left_idx = (p - 1) - side_right_idx
+
+    hit_turn = 5.0
+    wall_gain = 0.40
+    wall_target = 0.65
+
+
+    n_probe_self = 650.0
+    n_probe_neg = 650.0
+
+
+    t_bp = 0.003
+    g_refill = 650.0
+    g_drop = 220.0
+    g_hit = 0.8
+    g_side = 0.8
+    thr_u = 0.02
+
+
+    alpha_q = 0.18
+    beta_sigma = 0.12
+    k_uncertainty = 0.55
+    theta_commit = 0.02
+    choose_gain = 180.0
+
+    gate_switch = 1.2
+    gate_choose = 1.2
+
+
+
+    U_SELF_HIT = 0
+    U_SELF_HI = 1
+    U_SELF_LO = 2
+    U_NEG_HIT = 3
+    U_NEG_HI = 4
+    U_NEG_LO = 5
+
+    U_E = 20
+    U_E_PREV = 21
+    U_RISE = 22
+    U_DROP = 23
+    U_F1 = 24
+    U_BP = 25
+
+    U_MODE_NEG = 30
+    U_MODE_NEG_OVER = 31
+    U_SELF_ON = 32
+    U_NEG_ON = 33
+
+    U_COUNT_SELF = 40
+    U_COUNT_NEG = 41
+    U_SELF_DONE_RAW = 42
+    U_SELF_DONE_OVER = 43
+    U_SELF_DONE = 44
+    U_SELF_DONE_SEEN = 45
+    U_SELF_DONE_PULSE = 46
+    U_NEG_DONE_RAW = 47
+    U_NEG_DONE_OVER = 48
+    U_NEG_DONE = 49
+    U_NEG_DONE_SEEN = 50
+    U_NEG_DONE_PULSE = 51
+
+    U_U_SELF_RAW = 60
+    U_U_NEG_RAW = 61
+    U_U_SELF_ACT = 62
+    U_U_NEG_ACT = 63
+
+    U_Q_SELF = 70
+    U_Q_NEG = 71
+    U_ERR_SELF_POS = 72
+    U_ERR_SELF_NEG = 73
+    U_ERR_NEG_POS = 74
+    U_ERR_NEG_NEG = 75
+    U_SIGMA_SELF = 76
+    U_SIGMA_NEG = 77
+
+    U_A_RAW = 80
+    U_A_POS = 81
+    U_A_SCALE = 82
+    U_A_OVER = 83
+    U_A_IND = 84
+    U_CHOOSE_SELF_PULSE = 85
+
+    Win = np.zeros((n, 2 * p + 3), dtype=np.float64)
+    W = np.zeros((n, n), dtype=np.float64)
+    Wout = np.zeros((1, n), dtype=np.float64)
+
+
+
+    Win[U_E, energy_idx] = 1.0
+    W[U_E_PREV, U_E] = 1.0
+
+    Win[U_RISE, energy_idx] = 1.0
+    W[U_RISE, U_E] = -1.0
+
+    W[U_DROP, U_E] = 1.0
+    Win[U_DROP, energy_idx] = -1.0
+
+    W[U_F1, U_RISE] = 1.0
+    Win[U_F1, bias_idx] = -t_bp
+    W[U_BP, U_RISE] = 1.0
+    W[U_BP, U_F1] = -2.0
+
+
+
+    W[U_MODE_NEG, U_MODE_NEG] = 1.0
+    W[U_MODE_NEG, U_SELF_DONE_PULSE] = gate_switch
+    W[U_MODE_NEG, U_CHOOSE_SELF_PULSE] = -gate_choose
+    W[U_MODE_NEG, U_MODE_NEG_OVER] = -1.0
+
+    W[U_MODE_NEG_OVER, U_MODE_NEG] = 1.0
+    Win[U_MODE_NEG_OVER, bias_idx] = -1.0
+
+    Win[U_SELF_ON, bias_idx] = 1.0
+    W[U_SELF_ON, U_MODE_NEG] = -1.0
+    W[U_NEG_ON, U_MODE_NEG] = 1.0
+
+    W[U_COUNT_SELF, U_COUNT_SELF] = 1.0
+    W[U_COUNT_SELF, U_SELF_ON] = 1.0
+    W[U_COUNT_NEG, U_COUNT_NEG] = 1.0
+    W[U_COUNT_NEG, U_NEG_ON] = 1.0
+
+    W[U_SELF_DONE_RAW, U_COUNT_SELF] = 1.0
+    Win[U_SELF_DONE_RAW, bias_idx] = -n_probe_self
+    W[U_SELF_DONE_OVER, U_SELF_DONE_RAW] = 1.0
+    Win[U_SELF_DONE_OVER, bias_idx] = -1.0
+    W[U_SELF_DONE, U_SELF_DONE_RAW] = 1.0
+    W[U_SELF_DONE, U_SELF_DONE_OVER] = -1.0
+    W[U_SELF_DONE_SEEN, U_SELF_DONE_SEEN] = 1.0
+    W[U_SELF_DONE_SEEN, U_SELF_DONE] = 1.0
+    W[U_SELF_DONE_PULSE, U_SELF_DONE] = 1.0
+    W[U_SELF_DONE_PULSE, U_SELF_DONE_SEEN] = -1000.0
+
+    W[U_NEG_DONE_RAW, U_COUNT_NEG] = 1.0
+    Win[U_NEG_DONE_RAW, bias_idx] = -n_probe_neg
+    W[U_NEG_DONE_OVER, U_NEG_DONE_RAW] = 1.0
+    Win[U_NEG_DONE_OVER, bias_idx] = -1.0
+    W[U_NEG_DONE, U_NEG_DONE_RAW] = 1.0
+    W[U_NEG_DONE, U_NEG_DONE_OVER] = -1.0
+    W[U_NEG_DONE_SEEN, U_NEG_DONE_SEEN] = 1.0
+    W[U_NEG_DONE_SEEN, U_NEG_DONE] = 1.0
+    W[U_NEG_DONE_PULSE, U_NEG_DONE] = 1.0
+    W[U_NEG_DONE_PULSE, U_NEG_DONE_SEEN] = -1000.0
+
+
+
+
+    Win[U_U_SELF_RAW, side_right_idx] = g_side
+    W[U_U_SELF_RAW, U_BP] = g_refill
+    W[U_U_SELF_RAW, U_DROP] = -g_drop
+    Win[U_U_SELF_RAW, hit_idx] = -g_hit
+    Win[U_U_SELF_RAW, bias_idx] = -thr_u
+
+    Win[U_U_NEG_RAW, side_left_idx] = g_side
+    W[U_U_NEG_RAW, U_BP] = g_refill
+    W[U_U_NEG_RAW, U_DROP] = -g_drop
+    Win[U_U_NEG_RAW, hit_idx] = -g_hit
+    Win[U_U_NEG_RAW, bias_idx] = -thr_u
+
+
+    W[U_U_SELF_ACT, U_U_SELF_RAW] = 1.0
+    W[U_U_SELF_ACT, U_SELF_ON] = 1.0
+    Win[U_U_SELF_ACT, bias_idx] = -1.0
+
+    W[U_U_NEG_ACT, U_U_NEG_RAW] = 1.0
+    W[U_U_NEG_ACT, U_NEG_ON] = 1.0
+    Win[U_U_NEG_ACT, bias_idx] = -1.0
+
+
+    W[U_Q_SELF, U_Q_SELF] = 1.0 - alpha_q
+    W[U_Q_SELF, U_U_SELF_ACT] = alpha_q
+    W[U_Q_NEG, U_Q_NEG] = 1.0 - alpha_q
+    W[U_Q_NEG, U_U_NEG_ACT] = alpha_q
+
+
+    W[U_ERR_SELF_POS, U_U_SELF_ACT] = 1.0
+    W[U_ERR_SELF_POS, U_Q_SELF] = -1.0
+    W[U_ERR_SELF_NEG, U_Q_SELF] = 1.0
+    W[U_ERR_SELF_NEG, U_U_SELF_ACT] = -1.0
+
+    W[U_ERR_NEG_POS, U_U_NEG_ACT] = 1.0
+    W[U_ERR_NEG_POS, U_Q_NEG] = -1.0
+    W[U_ERR_NEG_NEG, U_Q_NEG] = 1.0
+    W[U_ERR_NEG_NEG, U_U_NEG_ACT] = -1.0
+
+
+    W[U_SIGMA_SELF, U_SIGMA_SELF] = 1.0 - beta_sigma
+    W[U_SIGMA_SELF, U_ERR_SELF_POS] = beta_sigma
+    W[U_SIGMA_SELF, U_ERR_SELF_NEG] = beta_sigma
+
+    W[U_SIGMA_NEG, U_SIGMA_NEG] = 1.0 - beta_sigma
+    W[U_SIGMA_NEG, U_ERR_NEG_POS] = beta_sigma
+    W[U_SIGMA_NEG, U_ERR_NEG_NEG] = beta_sigma
+
+
+    W[U_A_RAW, U_Q_SELF] = 1.0
+    W[U_A_RAW, U_Q_NEG] = -1.0
+    W[U_A_RAW, U_SIGMA_SELF] = -k_uncertainty
+    W[U_A_RAW, U_SIGMA_NEG] = -k_uncertainty
+    Win[U_A_RAW, bias_idx] = -theta_commit
+
+    W[U_A_POS, U_A_RAW] = 1.0
+    W[U_A_SCALE, U_A_POS] = choose_gain
+    W[U_A_OVER, U_A_SCALE] = 1.0
+    Win[U_A_OVER, bias_idx] = -1.0
+    W[U_A_IND, U_A_SCALE] = 1.0
+    W[U_A_IND, U_A_OVER] = -1.0
+
+
+    W[U_CHOOSE_SELF_PULSE, U_A_IND] = 1.0
+    W[U_CHOOSE_SELF_PULSE, U_NEG_DONE_PULSE] = 1.0
+    Win[U_CHOOSE_SELF_PULSE, bias_idx] = -1.0
+
+
+
+    Win[U_SELF_HIT, hit_idx] = 1.0
+    W[U_SELF_HIT, U_SELF_ON] = 1.0
+    Win[U_SELF_HIT, bias_idx] = -1.0
+
+    Win[U_SELF_HI, side_right_idx] = 1.0
+    W[U_SELF_HI, U_SELF_ON] = 1.0
+    Win[U_SELF_HI, bias_idx] = -(wall_target + 1.0)
+
+    Win[U_SELF_LO, side_right_idx] = -1.0
+    W[U_SELF_LO, U_SELF_ON] = 1.0
+    Win[U_SELF_LO, bias_idx] = wall_target - 1.0
+
+    Win[U_NEG_HIT, hit_idx] = 1.0
+    W[U_NEG_HIT, U_NEG_ON] = 1.0
+    Win[U_NEG_HIT, bias_idx] = -1.0
+
+    Win[U_NEG_HI, side_left_idx] = 1.0
+    W[U_NEG_HI, U_NEG_ON] = 1.0
+    Win[U_NEG_HI, bias_idx] = -(wall_target + 1.0)
+
+    Win[U_NEG_LO, side_left_idx] = -1.0
+    W[U_NEG_LO, U_NEG_ON] = 1.0
+    Win[U_NEG_LO, bias_idx] = wall_target - 1.0
+
+    Wout[0, U_SELF_HIT] = +hit_turn
+    Wout[0, U_SELF_HI] = +wall_gain
+    Wout[0, U_SELF_LO] = -wall_gain
+    Wout[0, U_NEG_HIT] = -hit_turn
+    Wout[0, U_NEG_HI] = -wall_gain
+    Wout[0, U_NEG_LO] = +wall_gain
+
+    f = relu
+    g = identity
+    yield Win, W, Wout, warmup, leak, f, g
+
+
+if __name__ == "__main__":
+    from pathlib import Path
+    import sys
+
+    root = Path(__file__).resolve().parents[2]
+    sys.path.insert(0, str(root / "braincraft" / "braincraft"))
+
+    from challenge_3 import train, evaluate
+    from bot import Bot
+    from environment_3 import Environment
+
+    np.random.seed(12345)
+    model = train(ensemble_counterfactual_twin_delib_v2_player, timeout=100.0)
+    score, std = evaluate(model, Bot, Environment, runs=10, seed=12345, debug=False)
+    print(f"score={score:.6f} std={std:.6f}")