Energy-Efficiency of Reinforcement Learning

Scope: Master’s project/Master’s thesis

Advisor: Dr. Kevin S. Luck (

Added: 31.August 2022

Topic: The advent of deep learning has brought forward a large range of offline and online reinforcement learning methodologies capable of learning complex tasks in robotics ranging from walking to manipulation. Many of these algorithms, especially in the case of deep reinforcement learning, are computationally expensive and require hours if not days of training time. This thesis will investigate the energy consumption of such reinforcement learning algorithms for both training and inference using monitoring capabilities. The goal is to find out how different algorithms compare in performance vs energy consumption on practical applications and if there are ways to reduce energy consumption by trading performance for low computational complexity, for example. Possible further extensions of the topic are possible, for example investigating the possibility to use energy-efficient reinforcement learning in combination with carbon-reduced policy behaviour, with potential applications to Autonomous Driving.

Minimum knowledge:

  • Python
  • Linux
  • Ability to perform literature surveys, understand and examine code bases/repositories

Preferred knowledge:

  • Comfortable with using PyTorch
  • Has taken or is taking classes in Machine Learning & Reinforcement Learning

Thesis student will:

  • Get an overview over different reinforcement leanring algorithms, their trade-offs and application areas
  • Learn to use reinforcement learning algorithms in practice and on relevant control tasks
  • Learn to use different reinforcement learning frameworks

Related Literature: