OPEN-SOURCE SCRIPT
AetherEdge - Self-Evolving RL Dashboard Lite
🖊️ Overview
A streamlined self-evolving agent built on a lightweight Actor-Critic core with prioritized experience, learning by pitting multiple specialist RL heads against each other. Internally, three heads — TREND-FOLLOWING, MEAN-REVERSION, and BREAKOUT — each propose a directional view from their own lens, while a meta-selector tracks each head's recent reward performance and re-allocates trust toward whichever is working now. As the market's character shifts, the agent automatically moves its weight toward the style that's paying off. Visualization is intentionally minimal: one main signal arrow, one prediction line, a small composite score, and a "Learning Progress" label that makes the evolution visible.
🔶 Key Features
Competing RL heads — trend / mean-reversion / breakout specialists run in parallel; the best drives the unified call
Dynamic trust re-allocation — softmax trust weights adapt to each head's reward-performance score
Lightweight Actor-Critic — each head has an Actor (policy) and Critic (state value), learning via advantage
Prioritized experience replay — high-reward transitions re-learned preferentially
Style-specific reward shaping — each head is rewarded more in its specialty regime, differentiating behavior
Ultra-minimal visuals — just one main arrow + one prediction line + a small composite score
Learning Progress label — visualizes agent maturity (bars trained, buffer fill, trust concentration) as a bar
Exploration vs. exploitation — ε-greedy exploration
🧠 Technical Architecture
The agent perceives the market as a four-dimensional state vector (momentum spread, RSI deviation, Bollinger position, position-within-range — all z-normalized). The core is three independent Actor-Critic heads, each producing Actor logits (3 actions → softmax policy) and a Critic value from a compact shared trunk (tanh).
Heads differentiate via style-specific reward shaping. From a common directional return (ATR-normalized), the trend head amplifies reward in strong trends (large |momentum|), the mean-reversion head when stretched from the mean (large |BB position|), and the breakout head near range edges. The three thus learn distinct behaviors from the same price data.
Learning is standard Actor-Critic. Each head's Critic computes the TD target r + γV(s′), the advantage A = TD target − V(s) drives the Actor's advantage-weighted policy gradient, and the Critic regresses via squared error — all from a prioritized replay buffer (|reward|^exponent). In meta-selection, each head's performance score is EMA-smoothed and turned into trust weights via softmax. The final policy blends the three heads' policies by trust and picks the max-value action. The composite score is the decisiveness of the blended policy; learning progress combines training maturity, buffer fill, and trust concentration.
⚙️ Recommended Settings & Tuning Guide
BTC (1H–4H): Training Horizon 800–1500, γ 0.93, Head Trust Adapt 0.05, Min Score 0.35. Standard settings fit well
ETH (1H–4H): As BTC, with Reward Lookahead 5–8
SOL (15m–1H): High volatility favors Head Trust Adapt ≈ 0.08 to switch styles quickly, Min Score 0.4 to be selective
XRP (1H–4H): Spike-prone; Priority Exponent ≈ 2.0 to learn hard from big experiences, shorter Prediction Length
Head Trust Adapt Rate: higher switches to the winning style faster but less stably; lower is stable but slower
Head Score Smoothing: higher (0.9–0.95) smooths performance assessment; lower is sensitive to recent results
Min Score to Signal: higher makes arrows more selective — tune to your trade frequency
Prediction Length: the projection distance of the line; a higher score steepens its slope
💡 How to Use in Practice
Reacting to the main arrow: an arrow is the "final verdict" where the trust-weighted consensus of all three heads clears the score gate — a core basis for trend-following entries
Using the prediction line: one line whose slope scales with score; steeper means a more confident directional view
Reading the composite score: a higher score % in the small label means clearer cross-head consensus; low means a balanced call (stand aside)
Using the Learning Progress label: as the "LEARNING" bar fills and the best head stabilizes, the policy is maturing; treat signals as indicative while progress is low
Watching the best head: the current leading head (TREND / MEAN-REV / BREAKOUT) in the progress label reveals how the agent reads the present market
Multi-timeframe usage: confirm the best head and direction on the higher timeframe (4H), then refine timing on arrows on the lower one (15m–1H)
⚠️ Important Notes
Initial learning period: right after launch the buffer is nearly empty and the heads are immature; treat signals as low-confidence until the Learning Progress bar fills (several hundred bars)
Learning resets: changing parameters, switching symbol/timeframe, or recompiling reinitializes all head weights, trust weights, and the buffer, restarting learning from zero
Nature of the Lite build: this distills the multi-head internals into ultra-simple output, omitting detailed internal-state visualization
On forward-looking reward: rewards use closed-bar forward returns (a standard RL training construct); current-bar decisions are on confirmed values, but as with any adaptive system, historical and live behavior can differ — always forward-test
Constraints: this is a lightweight implementation operating within Pine's compute budget
🚨 Disclaimer
This indicator is an analytical and educational visualization tool. The Actor-Critic, competing RL heads, prioritized experience replay, reward shaping, composite score, and prediction are quantitative heuristics computed on-chart from price data — they are not financial advice, buy/sell signals, or any guarantee of future performance. Reinforcement-learning agents can and do make wrong calls. Always combine any tool with your own analysis and disciplined risk management.
A streamlined self-evolving agent built on a lightweight Actor-Critic core with prioritized experience, learning by pitting multiple specialist RL heads against each other. Internally, three heads — TREND-FOLLOWING, MEAN-REVERSION, and BREAKOUT — each propose a directional view from their own lens, while a meta-selector tracks each head's recent reward performance and re-allocates trust toward whichever is working now. As the market's character shifts, the agent automatically moves its weight toward the style that's paying off. Visualization is intentionally minimal: one main signal arrow, one prediction line, a small composite score, and a "Learning Progress" label that makes the evolution visible.
🔶 Key Features
Competing RL heads — trend / mean-reversion / breakout specialists run in parallel; the best drives the unified call
Dynamic trust re-allocation — softmax trust weights adapt to each head's reward-performance score
Lightweight Actor-Critic — each head has an Actor (policy) and Critic (state value), learning via advantage
Prioritized experience replay — high-reward transitions re-learned preferentially
Style-specific reward shaping — each head is rewarded more in its specialty regime, differentiating behavior
Ultra-minimal visuals — just one main arrow + one prediction line + a small composite score
Learning Progress label — visualizes agent maturity (bars trained, buffer fill, trust concentration) as a bar
Exploration vs. exploitation — ε-greedy exploration
🧠 Technical Architecture
The agent perceives the market as a four-dimensional state vector (momentum spread, RSI deviation, Bollinger position, position-within-range — all z-normalized). The core is three independent Actor-Critic heads, each producing Actor logits (3 actions → softmax policy) and a Critic value from a compact shared trunk (tanh).
Heads differentiate via style-specific reward shaping. From a common directional return (ATR-normalized), the trend head amplifies reward in strong trends (large |momentum|), the mean-reversion head when stretched from the mean (large |BB position|), and the breakout head near range edges. The three thus learn distinct behaviors from the same price data.
Learning is standard Actor-Critic. Each head's Critic computes the TD target r + γV(s′), the advantage A = TD target − V(s) drives the Actor's advantage-weighted policy gradient, and the Critic regresses via squared error — all from a prioritized replay buffer (|reward|^exponent). In meta-selection, each head's performance score is EMA-smoothed and turned into trust weights via softmax. The final policy blends the three heads' policies by trust and picks the max-value action. The composite score is the decisiveness of the blended policy; learning progress combines training maturity, buffer fill, and trust concentration.
⚙️ Recommended Settings & Tuning Guide
BTC (1H–4H): Training Horizon 800–1500, γ 0.93, Head Trust Adapt 0.05, Min Score 0.35. Standard settings fit well
ETH (1H–4H): As BTC, with Reward Lookahead 5–8
SOL (15m–1H): High volatility favors Head Trust Adapt ≈ 0.08 to switch styles quickly, Min Score 0.4 to be selective
XRP (1H–4H): Spike-prone; Priority Exponent ≈ 2.0 to learn hard from big experiences, shorter Prediction Length
Head Trust Adapt Rate: higher switches to the winning style faster but less stably; lower is stable but slower
Head Score Smoothing: higher (0.9–0.95) smooths performance assessment; lower is sensitive to recent results
Min Score to Signal: higher makes arrows more selective — tune to your trade frequency
Prediction Length: the projection distance of the line; a higher score steepens its slope
💡 How to Use in Practice
Reacting to the main arrow: an arrow is the "final verdict" where the trust-weighted consensus of all three heads clears the score gate — a core basis for trend-following entries
Using the prediction line: one line whose slope scales with score; steeper means a more confident directional view
Reading the composite score: a higher score % in the small label means clearer cross-head consensus; low means a balanced call (stand aside)
Using the Learning Progress label: as the "LEARNING" bar fills and the best head stabilizes, the policy is maturing; treat signals as indicative while progress is low
Watching the best head: the current leading head (TREND / MEAN-REV / BREAKOUT) in the progress label reveals how the agent reads the present market
Multi-timeframe usage: confirm the best head and direction on the higher timeframe (4H), then refine timing on arrows on the lower one (15m–1H)
⚠️ Important Notes
Initial learning period: right after launch the buffer is nearly empty and the heads are immature; treat signals as low-confidence until the Learning Progress bar fills (several hundred bars)
Learning resets: changing parameters, switching symbol/timeframe, or recompiling reinitializes all head weights, trust weights, and the buffer, restarting learning from zero
Nature of the Lite build: this distills the multi-head internals into ultra-simple output, omitting detailed internal-state visualization
On forward-looking reward: rewards use closed-bar forward returns (a standard RL training construct); current-bar decisions are on confirmed values, but as with any adaptive system, historical and live behavior can differ — always forward-test
Constraints: this is a lightweight implementation operating within Pine's compute budget
🚨 Disclaimer
This indicator is an analytical and educational visualization tool. The Actor-Critic, competing RL heads, prioritized experience replay, reward shaping, composite score, and prediction are quantitative heuristics computed on-chart from price data — they are not financial advice, buy/sell signals, or any guarantee of future performance. Reinforcement-learning agents can and do make wrong calls. Always combine any tool with your own analysis and disciplined risk management.
Skrip sumber terbuka
Dalam semangat TradingView sebenar, pencipta skrip ini telah menjadikannya sumber terbuka, jadi pedagang boleh menilai dan mengesahkan kefungsiannya. Terima kasih kepada penulis! Walaupuan anda boleh menggunakan secara percuma, ingat bahawa penerbitan semula kod ini tertakluk kepada Peraturan Dalaman.
Penafian
Maklumat dan penerbitan adalah tidak bertujuan, dan tidak membentuk, nasihat atau cadangan kewangan, pelaburan, dagangan atau jenis lain yang diberikan atau disahkan oleh TradingView. Baca lebih dalam Terma Penggunaan.
Skrip sumber terbuka
Dalam semangat TradingView sebenar, pencipta skrip ini telah menjadikannya sumber terbuka, jadi pedagang boleh menilai dan mengesahkan kefungsiannya. Terima kasih kepada penulis! Walaupuan anda boleh menggunakan secara percuma, ingat bahawa penerbitan semula kod ini tertakluk kepada Peraturan Dalaman.
Penafian
Maklumat dan penerbitan adalah tidak bertujuan, dan tidak membentuk, nasihat atau cadangan kewangan, pelaburan, dagangan atau jenis lain yang diberikan atau disahkan oleh TradingView. Baca lebih dalam Terma Penggunaan.