Multi-Action TPG Algorithm for Multi-Task Reinforcement Learning
A novel benchmark utilizing the MuJoCo physics simulator has been established for Multi-Task Reinforcement Learning (MTRL) focusing on continuous control. The Multi-Action Tangled Program Graph (MATPG) algorithm, which is an adaptation of the Tangled Program Graph (TPG), combines MAPLE agents and formulates a control flow to engage them. MATPG was first evaluated in single-task RL environments, yielding results comparable to those of MAPLE. This research expands MATPG's application to multi-task settings, thereby offering a fresh benchmark for tasks requiring continuous control.
Key facts
- MATPG is a variation of the TPG algorithm.
- MATPG aggregates MAPLE agents.
- MATPG creates a control flow to activate agents.
- Initially tested on single-task RL environments.
- MATPG achieved similar results to MAPLE.
- A new benchmark based on MuJoCo is presented.
- The benchmark is for multi-task continuous control.
- The work is published on arXiv.
Entities
Institutions
- arXiv