README.md

May 16, 2025 · View on GitHub

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

This repository contains the official implementation of our paper "Many Minds, One Goal: Time Series Forecasting via Sub-task Specialization and Inter-agent Cooperation."

Project Structure

The project is divided into two subfolders according to the sub-task division strategies for agent specialization:

  • heterogeneous_sub_task: Agents are assigned heterogeneous forecasting tasks, each focusing on different aspects of the time series, including:

    • Statistical characteristics
    • Spectral energy distribution
    • Future seasonality
    • Future trend

  • homogeneous_sub_task: Agents are assigned homogeneous forecasting tasks, all predicting the same variable but at different temporal resolutions (i.e., multi-scale future forecasting).

Model Architecture

Each agent is implemented by iTransformer Architecture. The overall framework follows a pre-training → fine-tuning paradigm:

  • During pre-training, each agent is trained to specialize in its assigned forecasting sub-task. Agents communicate via the communication module using one of four predefined fixed-topology graphs:

    • Ring, Star, Chain, or Fully-connected These structures facilitate inter-agent information exchange—balancing specialization with global awareness.

  • In the fine-tuning phase, all agents are frozen, and the fixed communication edges are replaced with learnable weights, enabling the system to learn optimal communication flows specific to the forecasting objective.

  • A final Agent-Rated Voting Aggregator integrates the predictions from all agents to produce the final output, leveraging collective intelligence.

Datasets

We evaluate our method on 11 public benchmark datasets:

ETTh1, ETTh2, ETTm1, ETTm2, Weather, AQShunyi, AQWan, CzeLan, ZafNoo, PM2.5, Temp

Place all datasets under the ./dataset directory.

Running the Code

We provide all training scripts under ./scripts/{dataset_name}.sh. Common settings include:

--learning_rate 1e-3
--e_layers 2
--d_model 128
--d_ff 128

For single-agent comparisons, you may run experiments in the ./single_agent_forecasting folder under a fair setting using the same configuration.

Results

Our multi-agent forecasting system (MAFS) achieves a 6.35% improvement in forecasting performance compared to single-model baselines.

Despite using iTransformer (not the current SOTA) as the agent backbone, MAFS achieves:

  • 16× Top-1 performance
  • 4× Top-2 performance across 22 metrics on 11 datasets.

Ablation Studies

We further analyze the effect of the number of agents and the communication topology on performance:

Case Study