Barrier function inspired reward shaping in reinforcement learning

Includes bibliography.

MSc(Res);2024;Computer Science and Automation.

Reinforcement Learning (RL) has progressed from simple control tasks to complex real-world challenges with large state spaces. During initial iterations of training in most Reinforcement Learning (RL) algorithms, agents perform a significant number of random exploratory steps, which in the real world limits the practicality of these algorithms as this can lead to potentially dangerous behaviour. Hence, safe exploration is a critical issue when applying RL algorithms in the real world. Although RL excels in solving these challenging problems, the time required for convergence during training remains a significant limitation. Various techniques have been proposed to mitigate this issue, and reward shaping has emerged as a popular solution. However, most existing reward-shaping methods rely on value functions, which can pose scalability challenges as the environment’s complexity grows. Our research proposes a novel framework for reward shaping inspired by Barrier Functions, which is safety-oriented, intuitive, and easy to implement for any environment or task. To evaluate the effectiveness of our proposed reward formulations, we present our results on a challenging Safe Reinforcement Learning benchmark - the Open AI Safety Gym. We have conducted experiments on various environments, including CartPole, Half-Cheetah, Ant, and Humanoid. Our results demonstrate that our method leads to 1.4-2.8 times faster convergence and as low as 50-60% actuation effort compared to the vanilla reward. Moreover, our formulation has a theoretical basis for safety, which is crucial for real-world applications. In a sim-to-real experiment with the Go1 robot, we demonstrated better control and dynamics of the bot with our reward framework.