Many Minds, One Goal: Time Series Forecasting via Sub-task Specialization and Inter-agent Cooperation

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MAFS (Multi-Agent Forecasting System) replaces monolithic time series forecasting models with a system of specialized agents, each trained on a distinct temporal sub-task (multi-scale horizons or heterogeneous signal features), that communicate via graph convolution and aggregate predictions through a two-stage voting mechanism. Evaluated on 11 real-world benchmarks, it consistently outperforms single-model SOTA. The framework is architecture-agnostic and pluggable onto existing encoder backbones.

Problem

Monolithic forecasting models use a single architecture to handle all temporal patterns, domains, and horizons simultaneously, leading to poor generalization on heterogeneous or non-stationary time series. Mixture-of-Experts approaches partially address this but lack explicit inter-module communication and agent-level interpretability.

Method

MAFS follows a two-stage training paradigm:

Stage 1 – Specialization Pretraining: Each of N agents (iTransformer backbone) is assigned one of two sub-task decomposition strategies: (i) homogeneous — each agent forecasts a different future horizon prefix (multi-scale); (ii) heterogeneous — agents are assigned complementary signal analysis tasks (trend, seasonality, spectral energy, statistical moments). Agents communicate at every encoder layer via GCN-based message passing over a fixed topology (ring/star/chain/fully-connected) with uniform edge weights; only agent parameters are optimized.

Stage 2 – Collaborative Forecasting: Agent parameters are frozen. A Topology-aware Weight Adjuster makes edge weights learnable (soft adjacency = sigmoid(E) ⊙ fixed mask), allowing task-specific communication strengths to emerge. An Agent-rated Voting Aggregator (AVA) then integrates outputs: an Agent Confidence Estimator gates each agent's final representation against a shared global context vector, and a Global Voter computes MLP-derived per-agent collaboration weights over the gated representations, producing a unified latent that is linearly projected to the forecast.

Key Contributions

• First general-purpose MAS framework for time series forecasting (MAFS), establishing principled sub-task decomposition for temporal prediction.
• Two complementary task decomposition strategies (homogeneous multi-scale horizons; heterogeneous signal characteristics) enabling agent specialization.
• GCN-based inter-agent communication over configurable topologies with learnable edge weights in stage 2.
• Two-stage Agent-rated Voting Aggregator (confidence estimation + global voting) for adaptive ensemble fusion without agent collapse.
• Extensive ablation quantifying each module's contribution across 11 datasets.

Results

• MAFS ranks 1st on 16/22 metrics and top-2 on 20/22 metrics (MSE + MAE across 11 datasets).
• +6.35% MSE reduction, +4.03% MAE reduction vs. single-agent iTransformer baseline.
• +3.24% MSE, +1.71% MAE vs. TimeMixer (prior SOTA).
• Largest gain: Temp dataset, −19.08% MSE, −10.28% MAE over single-agent.
• Ablation (averaged over 11 datasets): removing AVA hurts most (−7.24% MSE, −5.70% MAE); removing sub-task specialization (−6.97% MSE, −3.93% MAE); removing communication (−4.18% MSE, −2.62% MAE).
• Star topology outperforms ring, chain, and fully-connected; performance saturates at ~16 agents.

Limitations

• Backbone is fixed to iTransformer; generalization to other backbone families is claimed but not empirically demonstrated across architectures in the main text.
• Agent count and topology must be hand-selected per dataset; no automatic configuration strategy is provided.
• Two-stage training requires twice the training overhead and separate hyperparameter tuning for each stage.
• Homogeneous vs. heterogeneous decomposition choice depends on dataset characteristics (stable vs. variable patterns) with no automatic selection criterion.
• LLM-based agents are not used; the "multi-agent" framing is architectural rather than involving language models or emergent reasoning.

Relevance to Agentic AI / LLM Agents

MAFS operationalizes core multi-agent principles — specialization, structured communication, and voting-based consensus — in a non-LLM forecasting setting, providing a clean empirical testbed for these design choices. The two-stage train-then-coordinate paradigm (freeze specialists, learn communication weights) mirrors patterns discussed in LLM multi-agent literature around role assignment and emergent collaboration. For researchers tracking agentic AI, this work offers concrete ablation evidence on which coordination mechanisms (aggregation > specialization > communication) matter most, and demonstrates that task decomposition granularity and topology choice are first-class architectural decisions rather than implementation details. It also highlights a gap: current LLM-agent frameworks rarely optimize inter-agent communication weights end-to-end, which MAFS achieves via differentiable graph topology learning.