Multi-Task Learning for Few-Shot Online Adaptation under Signal Temporal Logic Specifications

TL;DR

This paper presents a multi-task learning control framework using Signal Temporal Logic to efficiently adapt to new tasks with minimal data, demonstrating robustness in dynamical systems under perturbations.

Contribution

It introduces a novel MTL-based control approach with STL specifications, including a new methodology for learning and testing stages with robust ensemble initialization.

Findings

Achieves task compliance with few-shot learning in dynamical systems.

Demonstrates robustness under high perturbations.

Provides a new MTL framework for STL-based control.

Abstract

Multi-task learning (MTL) seeks to improve the generalized performance of learning specific tasks, exploiting useful information incorporated in related tasks. As a promising area, this paper studies an MTL-based control approach considering Signal Temporal Logic (STL). Task compliance is measured via the Robustness Degree (RD) which is computed by using the STL semantics. A suitable methodology is provided to solve the learning and testing stages, with an appropriate treatment of the non-convex terms in the quadratic objective function and using Sequential Convex Programming based on trust region update. In the learning stage, an ensemble of tasks is generated from deterministic goals to obtain a strong initializer for the testing stage, where related tasks are solved with a larger impact of perturbation. The methodology demonstrates to be robust in two dynamical systems showing results that meet the task specifications in a few shots for the testing stage, even for highly perturbed tasks.