EARCP: Self-Regulating Coherence-Aware Ensemble Architecture for Sequential Decision Making -- Ensemble Auto-Regule par Coherence et Performance

This paper introduces EARCP, a novel self-regulating ensemble architecture that dynamically weights heterogeneous expert models by balancing individual performance and inter-model coherence through an online learning mechanism, offering theoretical regret guarantees and demonstrating robustness in non-stationary sequential decision-making tasks.

The Gist

Imagine you are the captain of a ship navigating through a stormy, ever-changing ocean. You have a crew of four different navigators, each with their own unique style:

1. The Historian: Great at reading old maps and spotting patterns from the past.
2. The Meteorologist: Excellent at reading the wind and clouds right now.
3. The Astronomer: Good at looking far ahead at the stars for long-term direction.
4. The Mechanic: Knows exactly how the ship's engine is behaving and can predict mechanical failures.

The Problem with Traditional Teams

In most traditional "ensemble" teams (groups of experts), the captain picks a fixed rule: "I will listen to the Historian 25%, the Meteorologist 25%, and so on."

This works fine if the weather stays the same. But what if the Historian's old maps become useless because the coastline has changed? Or what if the Meteorologist is great at rain but terrible at fog? A fixed rule fails because it can't adapt when the world changes.

The Solution: EARCP (The "Smart Captain")

The paper introduces EARCP, a new way to run this team. Instead of a fixed rule, EARCP is a self-regulating, smart captain that constantly adjusts how much it listens to each navigator based on two things:

1. How well they are doing right now (Performance).
2. Whether they agree with each other (Coherence).

Here is how it works in simple terms:

1. The Scorecard (Performance)

Every time the ship makes a turn, the captain checks: "Did the Historian's prediction match where we actually ended up?"

• If the Historian was right, their score goes up.
• If they were wrong, their score goes down.
• This is like a teacher grading a student on a test.

2. The Group Chat (Coherence)

This is the secret sauce. EARCP doesn't just look at who is right; it looks at who agrees with whom.

• Imagine the Historian says, "Turn left!" and the Meteorologist says, "Turn left!" but the Astronomer says, "Turn right!"
• EARCP sees that two experts agree. It thinks, "Okay, even if I'm not 100% sure who is right, the fact that two of them agree gives me confidence."
• If everyone is shouting different things, the captain knows it's a "foggy" situation and becomes more cautious, perhaps listening to everyone a little bit more equally to avoid disaster.

3. The "Floor" Rule (Exploration)

Sometimes, a navigator might be having a bad day and get a low score. In a normal system, the captain might stop listening to them entirely.

• EARCP has a safety rule: "No one gets fired completely."
• It ensures every navigator gets at least a tiny bit of attention (a "floor"). Why? Because the Historian might be bad today but could be the only one who knows how to navigate a sudden storm tomorrow. This keeps the team ready for surprises.

Why is this better than the old ways?

• Old Way (Static): Like a robot that follows a script. If the script says "Trust the Historian," it trusts them even when the Historian is wrong.
• Old Way (Online Learning): Like a student who only cares about their own test score. They might ignore the fact that the whole class is confused, leading to bad decisions.
• EARCP: Like a wise leader who says, "You did well yesterday, but today you're struggling. Also, you and the Meteorologist agree, so I'll trust you a bit more. But I won't ignore the Astronomer completely, just in case."

Where can we use this?

The paper suggests this "Smart Captain" logic works anywhere the world changes quickly and we need to make decisions:

• Stock Markets: Instead of just trusting one trading algorithm, EARCP balances them. If the "Tech Stock" algorithm starts failing but agrees with the "Crypto" algorithm, it might adjust weights to handle the volatility.
• Medical Diagnosis: Imagine a system with an AI for X-rays, an AI for MRIs, and an AI for blood tests. If the X-ray AI is confused but the MRI and Blood Test AIs agree on a diagnosis, EARCP leans on that consensus.
• Self-Driving Cars: If the camera sensor is blinded by rain but the radar and GPS agree on the path, the car trusts the agreement rather than panicking.

The Bottom Line

EARCP is a framework that builds a team of AI experts that learns how to work together in real-time. It doesn't just ask, "Who is the smartest?" It asks, "Who is doing well right now, and who is agreeing with the group?"

By balancing individual performance with group agreement, and ensuring no expert is ever completely ignored, it creates a decision-making system that is robust, adaptable, and much harder to fool than a single model or a rigid team.

Technical Summary

Technical Summary: EARCP (Ensemble Auto-Régulé par Cohérence et Performance)

1. Problem Statement

The paper addresses critical limitations in traditional ensemble learning for sequential decision-making tasks. While ensemble methods generally outperform single models by leveraging diversity, standard approaches (e.g., static stacking, fixed-weight averaging) fail in dynamic environments due to three main challenges:

• Non-stationarity: Data distributions evolve over time, causing previously reliable models to degrade while others improve. Static weights cannot adapt to these shifts.
• Heterogeneity: Different model architectures (e.g., CNNs, LSTMs, Transformers) possess complementary strengths. Traditional methods often fail to optimally leverage this diversity in real-time.
• Partial/Delayed Feedback: In many applications (e.g., finance, robotics), the ground truth for a prediction is revealed with a significant delay, complicating the weight adaptation process.

Existing solutions like Mixture of Experts (MoE) often require joint offline training, while online algorithms like Hedge provide theoretical guarantees but ignore inter-model relationships (coherence) that could enhance robustness.

2. Methodology: The EARCP Framework

EARCP (Ensemble Auto-Régulé par Cohérence et Performance) is a novel online learning architecture that dynamically weights heterogeneous expert models. It balances exploitation (relying on high-performing models) with exploration (guided by consensus signals) through a principled multiplicative weight update mechanism.

Core Components

1. Dual-Statistic Tracking:

   • Performance Score ($P_{i,t}$): An Exponential Moving Average (EMA) of negative losses for each expert, capturing individual reliability.
   • Coherence Score ($C_{i,t}$): A measure of agreement between an expert and the rest of the ensemble.
     • Classification: Based on pairwise class agreement.
     • Regression: Based on inverse distance or correlation between predictions.
     • This score is also smoothed via EMA to reduce noise.

2. Weight Update Mechanism:
   The algorithm computes a combined score ($s_{i,t}$) for each expert $i$ at time $t$:
   $$s_{i,t} = \beta \cdot \tilde{P}_{i,t} + (1 - \beta) \cdot \tilde{C}_{i,t}$$
   Where:

   • $\beta \in [0, 1]$ is a hyperparameter balancing performance vs. coherence.
   • $\tilde{P}$ and $\tilde{C}$ are normalized scores.
   • The final weights are derived via an exponential transformation: $w_{i,t} \propto \exp(\eta_s \cdot s_{i,t})$.

3. Stabilization Techniques:

   • Floor Constraints: A minimum weight ($w_{min}$) is enforced to prevent weight collapse to a single expert, ensuring continued exploration.
   • Clipping & Normalization: Scores are clipped to prevent numerical overflow, and weights are renormalized to sum to 1.

Algorithm Flow

At each time step $t$:

1. Experts generate predictions.
2. The ensemble produces a weighted prediction.
3. Upon receiving the target (possibly delayed), losses are calculated.
4. Performance and Coherence scores are updated.
5. Weights are recalculated using the multiplicative update rule.

3. Key Contributions

1. Unified Framework: Introduces a formal method combining performance-based adaptation with coherence-aware weighting, enabling dynamic exploration and exploitation.
2. Theoretical Guarantees: Proves that EARCP achieves sublinear regret bounds of $O(\sqrt{T \log M})$.
   • When $\beta=1$ (pure performance), it matches the optimal bounds of the Hedge algorithm.
   • When $\beta < 1$, the regret bound increases by a factor of $1/\beta$, showing that coherence acts as side information without degrading worst-case guarantees significantly.
3. Practical Robustness: Implements stabilization techniques (floor constraints, EMA smoothing) specifically designed for non-stationary environments and delayed feedback.
4. Open-Source Implementation: Provides a complete Python library with experimental code to ensure reproducibility.

4. Experimental Results

The authors evaluated EARCP on three distinct sequential prediction domains: Electricity Consumption Forecasting, Human Activity Recognition (HAR), and Financial Time Series.

Performance Comparison

EARCP was compared against:

• Best Single Expert (Oracle)
• Equal Weighting
• Stacking (Offline meta-learner)
• Offline Mixture of Experts (MoE)
• Hedge Algorithm (Pure performance, no coherence)

Key Findings:

• Superior Accuracy: EARCP consistently outperformed all baselines with statistical significance ($p < 0.01$).
  • Electricity: 8.4% lower RMSE than the Hedge algorithm.
  • HAR: 3.8% higher accuracy than Offline MoE.
  • Finance: 10.5% better Sharpe ratio than Hedge.
• Robustness to Shifts: EARCP demonstrated superior adaptation during regime changes, maintaining stable performance while static ensembles degraded.
• Ablation Studies:
  • Removing coherence ($\beta=1$) degraded performance by 5–8%.
  • Removing the weight floor ($w_{min}=0$) caused weight collapse, reducing robustness.
• Efficiency: The computational overhead was minimal (<2ms per step beyond expert inference), dominated by expert predictions rather than the ensemble logic.

5. Significance and Applications

EARCP represents a significant advancement in online ensemble learning by formally integrating inter-model relationships (coherence) into the weight update process. Its significance lies in:

• Generalizability: The framework is domain-agnostic, applicable to any task with temporal dependencies and multiple predictive models.
• Real-World Applicability:
  • NLP: Dynamically weighting Large Language Models (LLMs) based on query type and consensus.
  • Medical Diagnosis: Aggregating multi-modal imaging models (CT, MRI) while tracking reliability over time.
  • Autonomous Systems: Robustly fusing perception and planning modules in robotics under changing environmental conditions.
  • Finance: Adapting to market regime shifts where model reliability fluctuates rapidly.

The paper concludes that EARCP offers a theoretically sound, computationally efficient, and practically robust solution for sequential decision-making in non-stationary environments, bridging the gap between theoretical online learning and practical ensemble deployment.