Deep Q-Network Signal Recommender

Interactive implementation of thesis research: Rekomendasi Channel Signal yang Dipersonalisasi Menggunakan Algoritma Reinforcement Learning. Features live Q-value neural updates, Nguyen-Widrow weight initialization, and real-time average diversity evaluation across multi-trader profiles.

Interactions (Experience)

94.0%

Recommendation Diversity

α = 0.2

Learning Rate & γ = 0.4

Nguyen-Widrow

Weight Scaler (β = 1.08)

📱 Trading Channel Feed (Live Re-ranking)

DQN Agent Output

Active Trader:

Channels are sorted dynamically by estimated Q-score. Subscribed channels receive penalty (-1.0) to move out of recommendations into dedicated subscribed lists, while views reinforce interest (+0.1).

Batch Experience Replay Simulation

🧠 DQN Matrix & Network State

Hidden Nodes: 15

Live Q-Value Matrix [User State × Channel Action]:

Thesis Formula 2.3 & Reward Policy (Table 3.1)

• DQN Update: Q(s, a) ← Q(s, a) + α [ r + γ max Q(s', a') - Q(s, a) ]

• Rewards: View = +0.1 | Subscribe = -1.0 | Unsubscribe = +0.2

• Activation: Bipolar Sigmoid f(x) = (2 / (1 + e^-x)) - 1

• Diversity: Diversity(c₁..c_n) = Σ (1 - Sim(c_i, c_j)) / (n(n-1)/2) → 94.0%

Reinforcement Engine Log Terminal

[SYS] Deep Q-Network Recommender initialized via Nguyen-Widrow weights.

[SYS] Hyperparameters: α = 0.2, γ = 0.4, Hidden Nodes = 15.

📑 Research Summary: PT Traders Family Signal Recommendation

Thesis 32160069

1. Nguyen-Widrow vs Random Initialization

Standard random initialization causes vanishing gradients when computing bipolar sigmoid activations across dense signal states. Applying the Nguyen-Widrow algorithm scales initial weight vectors (β = 0.7 × √ⁿp) to distribute intervals uniformly along the active activation curve, accelerating convergence by the 15th iteration without saturated weights.

2. Recommendation Diversity (94.0%)

Evaluated across 11,300 interaction records from 30 active trading accounts. Rather than collapsing into a popularity contest where all users receive the identical top channel, the DQN agent achieved an average similarity score of 6.0% (94.0% diversity), proving personalized exploration for scalping, swing trading, and commodity strategies.