Akshay Mete
Portrait of Akshay Mete

Akshay Mete

Postdoctoral Researcher, Electrical and Computer Engineering
Texas A&M University

I work on building AI agents that understand and act reliably in the physical world. Drawing on the long-standing theory of adaptive control, I address fundamental questions in world models, such as exploration and adaptation.

I am a postdoctoral researcher at Texas A&M University, working with Prof. P. R. Kumar, with whom I also completed my PhD in 2026. Before that, I completed a dual degree (B.Tech. + M.Tech.) in Electrical Engineering at IIT Bombay.

I am on the academic job market for 2026–27.

Research

Training objectives for deep world models

Optimistic World Models change how a world model is trained: an optimistic dynamics loss, derived from reward-biased maximum likelihood estimation in adaptive control, biases imagined transitions toward higher-reward outcomes. The loss is fully gradient-based, needs no uncertainty estimates, and plugs into existing world-model architectures; we instantiate it in DreamerV3 and STORM.

Key paper: arXiv 2026

Reward-biased model learning

Reward-Biased Maximum Likelihood Estimation (RBMLE) fits a model of the environment by maximum likelihood, biased toward models that promise higher reward, so that the model the agent learns is shaped by what it needs to control. I have developed RBMLE for Markov decision processes and linear systems, with applications to safe, multi-objective, and RLHF settings.

Key papers: L4DC 2021 · Allerton 2023

Learning models in closed loop, with guarantees

For linear systems, I study how well an agent can learn its model while controlling with it, measured by regret. PIECE is a self-tuning regulator for auto-regressive systems with the first O(log T) regret bound for self-tuning regulation, and augmented RBMLE-UCB brings reward-biased model learning to adaptive linear–quadratic control.

Key papers: ICML 2024 · NeurIPS 2022

Publications

World Models, Reinforcement Learning, and Control

Networks and Cyber-Physical Systems

Teaching

Fall 2026

ECEN 755: Stochastic Systems

Guest Lecturer, Texas A&M University

Lectures on dynamic programming: deterministic and stochastic DP, the Hamilton–Jacobi–Bellman equation, and infinite-horizon discounted problems.

2023

Reinforcement Learning: A Control Theoretic Perspective

Long tutorial, Indian Control Conference (ICC 2023)

Talks & Service

Invited Talks

  • From Adaptive Control to Reinforcement LearningSeminar, Purdue University

Reviewing

ICML · NeurIPS · AISTATS · AAAI · L4DC · CDC · ACC · ICC

Background

2026 – now

Postdoctoral Researcher, Electrical and Computer Engineering

Texas A&M University, College Station, TX

Host: P. R. Kumar

2019 – 2026

PhD, Electrical and Computer Engineering

Texas A&M University, College Station, TX

Dissertation: The Exploration–Exploitation Trade-Off in Reinforcement Learning
Advisor: P. R. Kumar

Summer 2024

Research Intern, AT&T Labs

Network Analytics and Automation Group, Bedminster, NJ

Reinforcement learning for network slicing. Mentor: H. Wang

2013 – 2018

B.Tech. + M.Tech. (Dual Degree), Electrical Engineering

Indian Institute of Technology Bombay, Mumbai

Communication and Signal Processing
Advisors: D. Manjunath, J. Nair, S. Moharir