Publications

Selected Publications

Adversarially Robust Multitask Adaptive Control

1) K. Fallah, L. F. Toso, J. Anderson (2025). Adversarially Robust Multitask Adaptive Control

^* = Alphabetical equal contribution.

Overview

The project studies how multiple control systems—each with uncertain or partially known dynamics—can collaboratively learn stabilizing controllers while remaining robust to heterogeneous models and adversarial (Byzantine) participants.

At the core of the method is a clustered multitask learning pipeline (illustrated below), which integrates

1. Data collection from multiple agents operating under certainty-equivalent controllers,
2. Clustered system identification (CSI) to group dynamically similar systems,
3. Resilient aggregation of model updates using (f, λ)-resilient rules such as the trimmed mean or geometric median, and
4. Controller synthesis via the discrete algebraic Riccati equation (DARE) for each identified cluster.

Figure 1: Workflow for adversarially robust multitask adaptive control.

This joint learning–control loop iteratively refines both the model estimates and the feedback gains, guaranteeing stable trajectories and sublinear cumulative regret.
Theoretical analysis shows that the expected regret for any honest system in cluster C_j scales as:

\[\mathbb{E}[R_T^{(i)}] = \tilde{\mathcal{O}}\left(\frac{\sqrt{dT}}{m_j}+ \lambda\sqrt{dT}+ \epsilon_{\mathrm{het}}^2 T\right)\]

Simulations

(a) Homogeneous clusters — collaboration reduces regret ≈ 1/√(cluster size).	(b) Heterogeneous clusters — bias floor ∝ εhet2·T.
(c) Misclassification rate — decays rapidly with more data per cluster.	(d) Adversarial ratios — sublinear regret sustained up to 30% corruption.
(e) Aggregators — trimmed mean vs geometric median (λ-resilient).	(f) Shared representation — RCSI remains robust without a global model.

Full details and codes can be found in this repository