MORE | Summer 2021
Learning Complex Behaviors from Simple Behaviors: An Analysis of Behavior-Based Modular Design for RL Agents
In this research, I present an algorithm that aims at combining behavior such that the resulting behavior has characteristics of each individual behavior. This approach has been inspired by subsumption architecture and motor schema-based mobile robot navigation in the domain of behavior-based robotics. The combination algorithm outputs n weights that combine n behaviors by a linear combination for combined behavior in the environment. Unlike traditional learning, the combination of behaviors is modular and can expedite learning by formulation of the right combination rather than updating the behaviors. This idea is tested on discrete and continuous environments like OpenAI’s “Lunar Lander” and “Biped Walker.” After experimentation, my analysis shows that the combination function is successfully able to combine behaviors. Future directions are also suggested as possible extensions to this research.
Kevin Jatin Vora
Hometown: Vadodara, Gujarat, India
Graduation date: Summer 2021