93c77b91b7
Mass deletion of the "v7 gem" reward terms identified in the Phase 1
inventory as category errors — each one rewarded an outcome-adjacent
behavior instead of encoding the underlying physics, and each
empirically hurt validation metrics more than it helped training
stability.
Deleted from experience_kernels.cu and all plumbing (Rust configs,
launch args, hyperopt logs, TOML entries):
* order_credit_weight - reward redundant with compute_tx_cost
(order_type_idx already differentiates
fills by order type)
* risk_efficiency_weight - reward double-counted drawdown penalty
asymmetrically (only on winners)
* urgency_credit_weight - reward was vol-normalized unrealized P&L,
pure rename of core return
* commitment_lambda - triple-counted churn + tx_cost
* w_dsr - kernel wrote DSR EMA but no longer added
to reward (dead); removed the EMA
bookkeeping too
* dsr_eta - kernel arg for the deleted DSR EMA
* position_entropy_weight - rewarded action-bucket diversity
regardless of outcome; histogram buffer
+ zero-init removed too
* exit_timing_weight - already inactive (used raw_next future
price, comment-deleted earlier)
* ofi_reward_weight - dead plumbing; OFI already passed as
feature through state[OFI_START..]
* opportunity_cost_scale - penalized flat when Q-gap wide;
redundant with Q-values themselves
Kernel arg count: experience_env_step_batch shrank from ~55 to ~45 args.
Rust-side config surface reduced correspondingly.
Results on E1 smoke test (20-epoch):
BEFORE any Phase 2 work:
Val Sharpe -120 to -150, MaxDD 10-15%, Sharpe_raw -0.39
AFTER reward_noise + Kelly (both envs) + urgency + this sweep:
Val Sharpe -17 to -22 (7× better)
Val MaxDD 0.27% (40× better)
Val Sharpe_raw ~-0.09 (4× better)
Training Sharpe_raw ~0 (stabilized from ±20 swings)
Final q_gap 0.1712 (highest yet, collapse mechanism fine)
The extreme train-Sharpe swings (+17 one epoch, -13 next) were not
learning dynamics — they were shaping-term noise. Core reward (P&L +
drawdown + churn + holding + tx_cost + Kelly physics cap) gives training
metrics that actually reflect what the model does.
Inventory doc (docs/superpowers/specs/2026-04-21-phase1-reward-inventory.md)
extended with a "better-form taxonomy" section: every deleted gem has
a correct layer it belongs to (physics, feature, diagnostic, gradient-
level regularization — not reward). Kelly cap and Q-target smoothing
are already relocated; others are scheduled per the taxonomy's P1/P2/P3
priority list.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
166 lines
4.3 KiB
TOML
166 lines
4.3 KiB
TOML
# DQN Hyperopt Profile — PSO search space definition
|
|
# All bounds are [min, max] ranges. The adapter reads these at runtime.
|
|
# To constrain the search space, narrow the ranges here — no code changes needed.
|
|
#
|
|
# Mean-reduced gradients (2026-04-09): LR range shifted down because gradient
|
|
# clipping no longer acts as a hidden LR reducer. Old range [1e-5, 3e-4] had
|
|
# effective LR of [~5e-10, ~1.5e-8] due to SUM-reduced clipping at batch=16384.
|
|
|
|
[search_space]
|
|
# Base parameters
|
|
learning_rate = [1e-7, 1e-4] # log scale in adapter — mean-reduced gradients
|
|
batch_size = [4096, 16384]
|
|
gamma = [0.90, 0.99] # wider range — v_range computed dynamically from gamma
|
|
buffer_size = [50000, 100000] # log scale in adapter
|
|
max_leverage = [2.0, 10.0] # leverage ratio; position computed from capital/price
|
|
huber_delta = [10.0, 40.0] # log scale in adapter
|
|
entropy_coefficient = [0.05, 0.5]
|
|
transaction_cost_multiplier = [0.5, 2.0]
|
|
per_alpha = [0.4, 0.8]
|
|
per_beta_start = [0.2, 0.6]
|
|
|
|
# Rainbow DQN extensions
|
|
v_range = [1.0, 1.0] # IGNORED: v_range now computed dynamically from gamma
|
|
noisy_sigma_init = [0.1, 1.0] # log scale in adapter
|
|
dueling_hidden_dim = [128, 512]
|
|
n_steps = [3, 5]
|
|
num_atoms = [52, 100] # 51=C51 paper minimum, 101=higher resolution (2x slower but better signal)
|
|
|
|
# Weight decay
|
|
weight_decay = [0.0001, 0.01] # log scale in adapter
|
|
|
|
# Kelly risk parameters
|
|
kelly_fractional = [0.25, 0.75]
|
|
kelly_max_fraction = [0.1, 0.5]
|
|
|
|
# Volatility
|
|
volatility_window = [10, 30]
|
|
|
|
# Soft update
|
|
tau = [0.005, 0.01] # log scale in adapter
|
|
|
|
# Network sizing
|
|
hidden_dim_base = [128, 256] # capped at production default: 512 is 4x slower for marginal benefit
|
|
|
|
# CQL regularization
|
|
cql_alpha = [0.5, 5.0]
|
|
|
|
# Training dynamics
|
|
lr_decay_type = [0, 2] # discrete: 0=constant, 1=linear, 2=cosine
|
|
minimum_profit_factor = [1.1, 2.0]
|
|
|
|
# Exploration
|
|
count_bonus_coefficient = [0.0, 0.3]
|
|
|
|
# Risk-adjusted returns
|
|
sharpe_weight = [0.0, 0.5]
|
|
|
|
# Branching DQN
|
|
branch_hidden_dim = [64, 256]
|
|
|
|
# Gradient accumulation
|
|
gradient_accumulation_steps = [1, 1] # fixed at 1 for now
|
|
|
|
# IQN dual-head
|
|
iqn_lambda = [0.0, 2.0]
|
|
|
|
# Spectral normalization
|
|
spectral_norm_sigma_max = [1.0, 10.0]
|
|
|
|
# C51 warmup
|
|
c51_warmup_epochs = [0, 10]
|
|
c51_alpha_max = [0.3, 0.9]
|
|
|
|
# HER ratio
|
|
her_ratio = [0.0, 0.8]
|
|
|
|
# Composite reward weights
|
|
w_pnl = [0.0, 1.0]
|
|
w_dd = [0.0, 5.0]
|
|
w_idle = [0.0, 0.1]
|
|
dd_threshold = [0.005, 0.03] # HFT: tight drawdown tolerance (0.5%-3%)
|
|
loss_aversion = [1.0, 1.0]
|
|
time_decay_rate = [0.0001, 0.005]
|
|
|
|
# Trade conviction filter
|
|
q_gap_threshold = [0.0, 0.5] # 0.0=trade every bar, 0.5=high conviction only
|
|
|
|
# Reward v7 — Counterfactual Branch Attribution
|
|
cea_weight = [0.1, 1.0]
|
|
# exposure_aux_weight removed (4-branch refactor)
|
|
b3_size = 3
|
|
|
|
# v8 search ranges
|
|
micro_reward_scale = [0.0, 0.05]
|
|
td_lambda = [0.5, 0.99]
|
|
hindsight_fraction = [0.0, 0.3]
|
|
hindsight_lookahead = [5, 20]
|
|
epsilon_start = [0.1, 0.5]
|
|
epsilon_end = [0.01, 0.05]
|
|
|
|
[experience]
|
|
initial_capital = 100000.0
|
|
|
|
[experience.fill_simulation]
|
|
ioc_fill_prob = 0.85
|
|
limit_fill_min = 0.30
|
|
limit_fill_max = 0.80
|
|
tx_cost_multiplier = 0.18 # IBKR ES: 0.18 bps = $4.50/contract RT
|
|
spread_cost_frac = 0.50
|
|
spread_capture_frac = 0.50
|
|
|
|
[risk]
|
|
q_clip_min = -200.0
|
|
q_clip_max = 200.0
|
|
|
|
[reward]
|
|
# v8 comprehensive training overhaul
|
|
micro_reward_scale = 0.01
|
|
td_lambda = 0.9
|
|
max_trace_length = 7
|
|
hindsight_fraction = 0.1
|
|
hindsight_lookahead = 10
|
|
w_pnl = 0.3
|
|
w_dd = 1.0
|
|
w_idle = 0.01
|
|
dd_threshold = 0.01
|
|
loss_aversion = 1.0
|
|
time_decay_rate = 0.0005
|
|
q_gap_threshold = 0.1
|
|
cea_weight = 0.3
|
|
b3_size = 3
|
|
|
|
[fixed]
|
|
|
|
[pso]
|
|
swarm_size = 20
|
|
max_iterations = 50
|
|
inertia = 0.7
|
|
cognitive = 1.5
|
|
social = 1.5
|
|
|
|
# Two-phase hyperopt: Phase 1 fixes architecture to small network,
|
|
# searches only learning dynamics (~15D). Phase 2 fixes best dynamics
|
|
# from Phase 1 JSON, searches architecture (~5D).
|
|
# See: docs/superpowers/specs/2026-03-22-two-phase-hyperopt-design.md
|
|
[phase_fast]
|
|
hidden_dim_base = 128
|
|
num_atoms = 101
|
|
branch_hidden_dim = 64
|
|
dueling_hidden_dim = 128
|
|
v_range = 1.0
|
|
# v8 phase_fast overrides
|
|
micro_reward_scale = 0.01
|
|
td_lambda = 0.9
|
|
hindsight_fraction = 0.1
|
|
hindsight_lookahead = 10
|
|
w_pnl = 0.3
|
|
w_dd = 1.0
|
|
w_idle = 0.01
|
|
dd_threshold = 0.01
|
|
loss_aversion = 1.0
|
|
time_decay_rate = 0.0005
|
|
q_gap_threshold = 0.1
|
|
cea_weight = 0.3
|
|
b3_size = 3
|