Multi-Task Optimization over Networks of Tasks
New algorithm scales to 5,000 tasks by treating them as graph nodes...
Researchers from multiple institutions have introduced MONET (Multi-Task Optimization over Networks of Tasks), a novel algorithm that rethinks how we tackle large-scale multi-task optimization. Traditional population-based methods struggle with scalability beyond a few hundred tasks, while MAP-Elites variants, though scalable, rely on fixed discretized archives that ignore the underlying topology of the task space. MONET addresses this by representing the task space as a graph—tasks form nodes, and edges connect tasks in the parameter space. This graph structure allows the algorithm to exploit task relationships for knowledge transfer while remaining computationally tractable for high-dimensional problems.
MONET operates through a dual learning mechanism: social learning generates candidate solutions by crossover between neighboring nodes on the graph, while individual learning refines each node's solution independently via mutation. The team evaluated MONET on four challenging domains—archery, arm, and cartpole (each with 5,000 tasks) and hexapod (2,000 tasks). Results show that MONET matches or exceeds the performance of existing MAP-Elites-based baselines across all four domains. This graph-based approach opens new possibilities for scaling multi-task optimization to thousands of tasks without sacrificing solution quality or requiring pre-defined task discretization.
- MONET models task space as a graph with nodes (tasks) and edges (parameter space connections), enabling topology-aware knowledge transfer
- Combines social learning (crossover from neighbors) and individual learning (mutation) for efficient optimization
- Matches or exceeds MAP-Elites baselines on 4 domains with up to 5,000 tasks each
Why It Matters
Graph-based task modeling enables scalable multi-task optimization for real-world problems with thousands of interconnected tasks.