# DQN Local Dev Profile — extended training on RTX 3050 (4GB VRAM)
# Larger buffer + more episodes than smoketest for meaningful convergence.
# gpu_n_episodes=64 × gpu_timesteps=200 = 12,800 experiences/epoch
# buffer_size=50K → fills over ~4 epochs, reducing memorization

[training]
epochs = 200
batch_size = 64
learning_rate = 0.0001
gamma = 0.95
weight_decay = 0.0001
adam_epsilon = 1e-8
warmup_steps = 0
hidden_dim_base = 64
max_steps_per_epoch = 200
reward_scale = 1.0
huber_delta = 1.0
lr_decay_type = 2
lr_min = 0.00001

[distributional]
num_atoms = 51

[exploration]
epsilon_start = 1.0
epsilon_end = 0.05
epsilon_decay = 0.85
noisy_sigma_init = 0.3
entropy_coefficient = 0.001
count_bonus_coefficient = 0.1
q_gap_threshold = 0.05

[replay_buffer]
buffer_size = 50000
min_replay_size = 500

[early_stopping]
enabled = true
patience = 20
min_epochs_before_stopping = 80

[experience]
gpu_n_episodes = 64
gpu_timesteps_per_episode = 200
min_hold_bars = 3

[experience.fill_simulation]
ioc_fill_prob = 0.85
limit_fill_min = 0.30
limit_fill_max = 0.80
spread_cost_frac = 0.50
spread_capture_frac = 0.50

[risk]
max_position = 2.0
loss_aversion = 1.5
q_clip_min = -50.0
q_clip_max = 50.0

[advanced]
n_steps = 3
tau = 0.005
c51_warmup_epochs = 5
her_ratio = 0.2
cql_alpha = 0.1
curiosity_weight = 0.1
iqn_lambda = 0.25
spectral_norm_sigma_max = 1.5
gradient_clip_norm = 1.0

[reward]
loss_aversion = 1.5
q_gap_threshold = 0.1