Safety and Life-long Reinforcement Learning
Random exploration is one of the main mechanisms through which reinforcement learning (RL) solves tasks. However, random exploration can lead to undesirable or catastrophic outcomes when learning in safety-critical environments. In fact, safe learning is one of the major obstacles towards real-world agents that can learn during deployment. This project revolves around approaches that automatically train sub-modules that shape learning such that the agent will instinctively avoid dangers while still being capable of exploring its surroundings to learn new tasks.
Grbic, D & Risi, S 2020, Safe Reinforcement Learning through Meta-learned Instincts. in ALIFE 2020: The 2020 Conference on Artificial Life. 32 edn, MIT Press, Artificial Life Conference Proceedings , pp. 183-291, ALIFE 2020, Montreal, Canada, 13/07/2020. https://doi.org/10.1162/isal_a_00318
Djordje Grbic, Sebastian Risi; July 19–23, 2021. “Safer Reinforcement Learning through Transferable Instinct Networks.” Proceedings of the ALIFE 2021: The 2021 Conference on Artificial Life. ALIFE 2021: The 2021 Conference on Artificial Life. Online. (pp. 107). ASME. https://doi.org/10.1162/isal_a_00449