Lecture Notes in Computer Science: Safe Exploration Techniques for Reinforcement Learning – an Overview

By M. Pecka, Tomáš Svoboda

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Reinforcement Learning

Citations in the corpus

Concrete Problems in AI Safety (2016, background): “[53, 116] provide thorough reviews of this literature, so we don’t review it extensively here, but simply describe some general routes that this research has taken, as well as suggesting some directions that might have increasing relevance as RL systems expand in scope and capability.”
Safe Model-based Reinforcement Learning with Stability Guarantees (2017, background): “Related work Safety is an active research topic in RL and different definitions of safety exist [5, 6].”
AI Safety Gridworlds (2017, background): “How can we build agents that respect the safety constraints not only during normal operation, but also during the initial learning period? This problem is known as the safe exploration problem (Pecka and Svoboda, 2014; Garcı́a and Fernández, 2015).”
Safe Exploration Algorithms for Reinforcement Learning Controllers (2018, background): “The goal of this paper is to avoid such occurrences during learning, thus achieving safe exploration [17].”
Safe Exploration in Finite Markov Decision Processes with Gaussian Processes (2016, background): “Safe exploration is an open problem in the reinforcement learning community and several definitions of safety have been proposed [16].”
Probabilistically safe policy transfer (2017, background): “A review of past work on safe exploration is given in [2], [12], [21].”
Measuring and avoiding side effects using relative reachability (2018, background): “This problem has mostly been studied in the context of safe exploration during the agent’s learning process (Pecka and Svoboda, 2014; García and Fernández, 2015), but it can also occur after training if the reward function does not incorporate disruptions to the environment.”
Safe Reinforcement Learning via Shielding (2017, background): “A number of different notions of safety were recently explored [8,15].”
Penalizing Side Effects using Stepwise Relative Reachability (2018, background): “For machine learning systems, it has mostly been studied in the context of safe exploration during the agent’s learning process [Pecka and Svoboda, 2014; García and Fernández, 2015], but can also occur after training if the reward function is misspecified and fails to penalize disruptions to the…”
Safely Interruptible Agents (2016, background): “On top of defining what is considered a good behaviour of the agent after learning, there may be physical safety constraints during learning [Pecka and Svoboda, 2014]: a robot should not harm its environment or break itself, in particular if it learns by trial and error like RL agents.”