Behavior adaptation from negative social feedback based on goal awareness
| Autor: | Philippe Gaussier, Antoine de Rengerve, Raphael Braud, Pierre Andry |
|---|---|
| Přispěvatelé: | De Rengervé, Antoine, Equipes Traitement de l'Information et Systèmes (ETIS - UMR 8051), Ecole Nationale Supérieure de l'Electronique et de ses Applications (ENSEA)-Centre National de la Recherche Scientifique (CNRS)-CY Cergy Paris Université (CY) |
| Jazyk: | angličtina |
| Rok vydání: | 2012 |
| Předmět: |
Social robot
business.industry 05 social sciences [INFO.INFO-NE] Computer Science [cs]/Neural and Evolutionary Computing [cs.NE] Context (language use) [INFO.INFO-LG] Computer Science [cs]/Machine Learning [cs.LG] [INFO.INFO-NE]Computer Science [cs]/Neural and Evolutionary Computing [cs.NE] Action selection Robot learning Human–robot interaction 03 medical and health sciences 0302 clinical medicine [INFO.INFO-LG]Computer Science [cs]/Machine Learning [cs.LG] Robot Latent learning 0501 psychology and cognitive sciences Artificial intelligence business Psychology Adaptation (computer science) 030217 neurology & neurosurgery 050104 developmental & child psychology |
| Zdroj: | ICDL-EPIROB 2012 IEEE International Conference on Development and Learning (ICDL)-Epigenetics and Robotics (Epirob) 2012 IEEE International Conference on Development and Learning (ICDL)-Epigenetics and Robotics (Epirob), Nov 2012, San Diego, CA, USA, United States. pp.1--6 |
| Popis: | International audience; Robots are expected to perform actions in a human environment where they will have to learn both how and when to act. Social human robot interaction could provide the robot with external feedback to guide them. In this paper, the focus is put on managing correctly negative signals thus stressing the importance of being aware of its own goal. In previous works, we developed bio-inspired models for action planning which enabled a robot to adapt its space representations and thus its behavior in the context of latent learning with rewards. Though, as the action selection is based on a local readout of a propagated gradient, the current goal is not explicitly available. To determine it, the implemented mechanisms are : first, to select and inhibit one of the potential goals and then, to monitor if this inhibition changes the current behavior of the agent. If so, the inhibited goal is the one pursued. As a result, negative signals can then be used to directly modulate the strength of the current goal and change the agent's behavior. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|