MONET: Graph-Based Multi-Task Optimization Algorithm
A new multi-task optimization algorithm called MONET (Multi-Task Optimization over Networks of Tasks) has been introduced. MONET models the task space as a graph where tasks are nodes and edges connect tasks in the parameter space, enabling knowledge transfer between tasks. It combines social learning (crossover from neighboring nodes) with individual learning (refinement of a node's own solution). The algorithm addresses limitations of existing methods: population-based approaches scale poorly for large task sets, and MAP-Elites variants rely on fixed discretized archives that ignore task space topology. MONET remains tractable for high-dimensional problems while exploiting topology. The work is described in arXiv preprint 2604.21991.
Key facts
- MONET stands for Multi-Task Optimization over Networks of Tasks.
- The task space is modeled as a graph with tasks as nodes and edges connecting tasks in parameter space.
- MONET combines social learning via crossover from neighboring nodes and individual learning.
- Existing population-based methods scale poorly for large task sets.
- MAP-Elites variants use fixed discretized archives that disregard task space topology.
- MONET is designed for high-dimensional problems.
- The algorithm enables knowledge transfer between tasks.
- The preprint is available on arXiv with ID 2604.21991.
Entities
Institutions
- arXiv