A social reinforcement learning agent
Proceedings of the fifth international conference on Autonomous agents
Introduction to Reinforcement Learning
Introduction to Reinforcement Learning
AAAI'06 Proceedings of the 21st national conference on Artificial intelligence - Volume 1
Interactively shaping agents via human reinforcement: the TAMER framework
Proceedings of the fifth international conference on Knowledge capture
Combining manual feedback with subsequent MDP reward signals for reinforcement learning
Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems: volume 1 - Volume 1
Reinforcement learning from simultaneous human and MDP reward
Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems - Volume 1
Machine learning for interactive systems and robots: a brief introduction
Proceedings of the 2nd Workshop on Machine Learning for Interactive Systems: Bridging the Gap Between Perception, Action and Communication
Hi-index | 0.00 |
In the last years there has been an increasing interest on using human feedback during robot operation to incorporate non-expert human expertise while learning complex tasks. Most work has considered reinforcement learning frameworks were human feedback, provided through multiple modalities (speech, graphical interfaces, gestures) is converted into a reward. This paper explores a different communication channel: cognitive EEG brain signals related to the perception of errors by humans. In particular, we consider error potentials (ErrP), voltage deflections appearing when a user perceives an error, either committed by herself or by an external machine, thus encoding binary information about how a robot is performing a task. Based on this potential, we propose an algorithm based on policy matching for inverse reinforcement learning to infer the user goal from brain signals. We present two cases of study involving a target reaching task in a grid world and using a real mobile robot, respectively. For discrete worlds, the results show that the robot is able to infer and reach the target using only error potentials as feedback elicited from human observation. Finally, promising preliminary results were obtained for continuous states and actions in real scenarios.