OPEN-SOURCE SCRIPT
AetherEdge - KNN-Augmented RL Hybrid
🖊️ Overview
A self-evolving hybrid that fuses episodic memory with a learned value network. Each bar's state and its realized outcome are stored in an experience memory bank; to value the present, the agent runs a k-Nearest-Neighbors search over that bank — retrieving the most similar historical states and distance-weighting their outcomes — then blends that memory estimate with a NeuraLib NN/Q estimate. The fusion combines the generalization of gradient learning with the instant recall of episodic memory. A reward-driven update adapts the per-feature similarity weights, so the agent learns which dimensions of "similarity" actually matter for prediction. Visualization: auto breakout-zone boxes where high-similarity memory predicts a decisive move, plus directional arrows on the fused signal.
🔶 Key Features
kNN×RL hybrid — episodic-memory kNN retrieval fused with a NeuraLib value network
Episodic memory augmentation — instant distance-weighted recall of similar past outcomes
Memory ↔ network blend — continuously adjustable from pure memory (kNN) to pure network (NN/Q)
Reward-driven similarity-weight learning — learns which feature dimensions define "similar"
High-similarity breakout zones — zones spawn only when memory retrieval is confident; opacity tracks similarity
Fused signal arrows — the integrated kNN+NN verdict shown as directional arrows (with glow)
Exploration vs. exploitation — ε-greedy exploration
Premium panel — NN-vs-kNN breakdown, fused value, memory similarity, conviction, similarity weights, memory/zone counts
🧠 Technical Architecture
The agent perceives the market as a four-dimensional state vector (momentum spread, RSI deviation, a volatility-regime ratio, position-within-range — all z-normalized). The architecture fuses two value-estimation paths.
Episodic memory (kNN) path: each bar stores state, action taken, and realized value (reward) in a memory bank. To value the present, a weighted Euclidean distance — using adaptive per-feature weights — is computed against all memories, and the k nearest are retrieved. Each neighbor's distance is turned into similarity via a Gaussian-style kernel sim = exp(−sharpness × distance), and a per-action similarity-weighted average yields the kNN memory value.
Value network (NeuraLib) path: a single hidden-layer (tanh) network maps state to per-action Q-values, trained by a DQN-style update.
Fusion: per direction, fused = (1−blend)·Q_network + blend·Q_knn, and the action with the largest fused value is selected. The signature element is similarity-weight adaptation: on each reward, weights of feature dimensions consistent with the profitable outcome are raised and irrelevant ones lowered (w ← clip(w + lr·(|feature|·|reward| − baseline))). The very definition of "similar" thus evolves from reward. Zone spawning fires only when both memory similarity and conviction clear their gates, with box opacity tracking similarity.
⚙️ Recommended Settings & Tuning Guide
BTC (1H–4H): Training Horizon 800–1500, Neighbors k 6, Memory Depth 300, Blend 0.5, γ 0.93. A 50/50 memory-network fusion as standard
ETH (1H–4H): As BTC, with Reward Lookahead 5–8
SOL (15m–1H): High volatility favors Similarity Sharpness 2.5–3.0 to weight only close neighbors, Min Similarity 0.7 to be selective on zones
XRP (1H–4H): Spike-prone; Blend ≈ 0.6 to lean on memory (recalling past sharp moves), k ≈ 8
Memory ↔ Network Blend: toward 0 favors generalization (robust to novel states); toward 1 favors memory (replays known patterns) — tune to market repetitiveness
Neighbors (k): small (3–5) is sharp but noisy; large (8–15) is smooth but less responsive
Similarity Weight LR: higher adapts feature weights faster but less stably; 0.01–0.03 is the stable zone
Memory Bank Depth: deeper retains more diverse memory but mixes in older regimes — balance against Training Horizon
💡 How to Use in Practice
Reacting to fused arrows: an arrow means both kNN memory and NN value back the direction and conviction cleared the gate — the integrated verdict, a core basis for trend-following entries
Using high-similarity zones: zones mark areas where memory shows a decisive move occurred in similar past conditions; watch breaks/bounces at zone edges
Reading memory similarity: higher panel similarity means the present closely matches a clear past pattern — higher prediction confidence; be cautious when low (novel conditions)
NN vs. kNN agreement: when both (e.g. the "NN / kNN" in the LONG row) are high together, confidence is highest; stand aside when they disagree
Watching similarity weights: the panel's Sim Weights (M·R·V·P = momentum, RSI, volatility, position) reveal which features the agent currently weights for similarity
Multi-timeframe usage: map high-similarity zones on the higher timeframe (4H), then refine timing on fused arrows on the lower one (15m–1H)
Combinations: filter for high-similarity, high-conviction arrows to focus on memory-backed, high-quality entries
⚠️ Important Notes
Initial learning period: right after launch the memory bank is nearly empty and kNN retrieval is non-functional; treat signals and zones as low-confidence until memory accumulates (several hundred bars)
Learning resets: changing parameters, switching symbol/timeframe, or recompiling reinitializes the memory bank, network weights, similarity weights, and zones, restarting learning from zero
Memory bias: the bank is built from recent states, so when the regime shifts substantially, past memory fits the present less well — low Memory Similarity is the tell
On forward-looking reward: rewards and memory values use closed-bar forward returns (a standard RL training construct); current-bar evaluation is 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; kNN search scales with memory depth × neighbors, so extreme settings affect performance
🚨 Disclaimer
This indicator is an analytical and educational visualization tool. The kNN search, value network, reward shaping, similarity-weight adaptation, breakout zones, and fused signal 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 self-evolving hybrid that fuses episodic memory with a learned value network. Each bar's state and its realized outcome are stored in an experience memory bank; to value the present, the agent runs a k-Nearest-Neighbors search over that bank — retrieving the most similar historical states and distance-weighting their outcomes — then blends that memory estimate with a NeuraLib NN/Q estimate. The fusion combines the generalization of gradient learning with the instant recall of episodic memory. A reward-driven update adapts the per-feature similarity weights, so the agent learns which dimensions of "similarity" actually matter for prediction. Visualization: auto breakout-zone boxes where high-similarity memory predicts a decisive move, plus directional arrows on the fused signal.
🔶 Key Features
kNN×RL hybrid — episodic-memory kNN retrieval fused with a NeuraLib value network
Episodic memory augmentation — instant distance-weighted recall of similar past outcomes
Memory ↔ network blend — continuously adjustable from pure memory (kNN) to pure network (NN/Q)
Reward-driven similarity-weight learning — learns which feature dimensions define "similar"
High-similarity breakout zones — zones spawn only when memory retrieval is confident; opacity tracks similarity
Fused signal arrows — the integrated kNN+NN verdict shown as directional arrows (with glow)
Exploration vs. exploitation — ε-greedy exploration
Premium panel — NN-vs-kNN breakdown, fused value, memory similarity, conviction, similarity weights, memory/zone counts
🧠 Technical Architecture
The agent perceives the market as a four-dimensional state vector (momentum spread, RSI deviation, a volatility-regime ratio, position-within-range — all z-normalized). The architecture fuses two value-estimation paths.
Episodic memory (kNN) path: each bar stores state, action taken, and realized value (reward) in a memory bank. To value the present, a weighted Euclidean distance — using adaptive per-feature weights — is computed against all memories, and the k nearest are retrieved. Each neighbor's distance is turned into similarity via a Gaussian-style kernel sim = exp(−sharpness × distance), and a per-action similarity-weighted average yields the kNN memory value.
Value network (NeuraLib) path: a single hidden-layer (tanh) network maps state to per-action Q-values, trained by a DQN-style update.
Fusion: per direction, fused = (1−blend)·Q_network + blend·Q_knn, and the action with the largest fused value is selected. The signature element is similarity-weight adaptation: on each reward, weights of feature dimensions consistent with the profitable outcome are raised and irrelevant ones lowered (w ← clip(w + lr·(|feature|·|reward| − baseline))). The very definition of "similar" thus evolves from reward. Zone spawning fires only when both memory similarity and conviction clear their gates, with box opacity tracking similarity.
⚙️ Recommended Settings & Tuning Guide
BTC (1H–4H): Training Horizon 800–1500, Neighbors k 6, Memory Depth 300, Blend 0.5, γ 0.93. A 50/50 memory-network fusion as standard
ETH (1H–4H): As BTC, with Reward Lookahead 5–8
SOL (15m–1H): High volatility favors Similarity Sharpness 2.5–3.0 to weight only close neighbors, Min Similarity 0.7 to be selective on zones
XRP (1H–4H): Spike-prone; Blend ≈ 0.6 to lean on memory (recalling past sharp moves), k ≈ 8
Memory ↔ Network Blend: toward 0 favors generalization (robust to novel states); toward 1 favors memory (replays known patterns) — tune to market repetitiveness
Neighbors (k): small (3–5) is sharp but noisy; large (8–15) is smooth but less responsive
Similarity Weight LR: higher adapts feature weights faster but less stably; 0.01–0.03 is the stable zone
Memory Bank Depth: deeper retains more diverse memory but mixes in older regimes — balance against Training Horizon
💡 How to Use in Practice
Reacting to fused arrows: an arrow means both kNN memory and NN value back the direction and conviction cleared the gate — the integrated verdict, a core basis for trend-following entries
Using high-similarity zones: zones mark areas where memory shows a decisive move occurred in similar past conditions; watch breaks/bounces at zone edges
Reading memory similarity: higher panel similarity means the present closely matches a clear past pattern — higher prediction confidence; be cautious when low (novel conditions)
NN vs. kNN agreement: when both (e.g. the "NN / kNN" in the LONG row) are high together, confidence is highest; stand aside when they disagree
Watching similarity weights: the panel's Sim Weights (M·R·V·P = momentum, RSI, volatility, position) reveal which features the agent currently weights for similarity
Multi-timeframe usage: map high-similarity zones on the higher timeframe (4H), then refine timing on fused arrows on the lower one (15m–1H)
Combinations: filter for high-similarity, high-conviction arrows to focus on memory-backed, high-quality entries
⚠️ Important Notes
Initial learning period: right after launch the memory bank is nearly empty and kNN retrieval is non-functional; treat signals and zones as low-confidence until memory accumulates (several hundred bars)
Learning resets: changing parameters, switching symbol/timeframe, or recompiling reinitializes the memory bank, network weights, similarity weights, and zones, restarting learning from zero
Memory bias: the bank is built from recent states, so when the regime shifts substantially, past memory fits the present less well — low Memory Similarity is the tell
On forward-looking reward: rewards and memory values use closed-bar forward returns (a standard RL training construct); current-bar evaluation is 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; kNN search scales with memory depth × neighbors, so extreme settings affect performance
🚨 Disclaimer
This indicator is an analytical and educational visualization tool. The kNN search, value network, reward shaping, similarity-weight adaptation, breakout zones, and fused signal 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.
Скрипт с открытым кодом
В истинном духе TradingView, создатель этого скрипта сделал его открытым исходным кодом, чтобы трейдеры могли проверить и убедиться в его функциональности. Браво автору! Вы можете использовать его бесплатно, но помните, что перепубликация кода подчиняется нашим Правилам поведения.
Отказ от ответственности
Информация и публикации не предназначены для предоставления и не являются финансовыми, инвестиционными, торговыми или другими видами советов или рекомендаций, предоставленных или одобренных TradingView. Подробнее читайте в Условиях использования.
Скрипт с открытым кодом
В истинном духе TradingView, создатель этого скрипта сделал его открытым исходным кодом, чтобы трейдеры могли проверить и убедиться в его функциональности. Браво автору! Вы можете использовать его бесплатно, но помните, что перепубликация кода подчиняется нашим Правилам поведения.
Отказ от ответственности
Информация и публикации не предназначены для предоставления и не являются финансовыми, инвестиционными, торговыми или другими видами советов или рекомендаций, предоставленных или одобренных TradingView. Подробнее читайте в Условиях использования.