Apprentissage adaptatif de comportements éthiques

Publication at a national conference (JFSMA'20).

Available online on HAL.

This paper (in French) presents a multi-agent Reinforcement Learning algorithm to learn "ethical behaviours", i.e., behaviours aligned with ethical considerations (moral values). It focuses on multi-dimensional and continuous situations and actions, and includes in-design ethical considerations, such as privacy, by avoiding to share individual data (observations, actions, rewards).

Abstract:

The increase in the use of Artificial Intelligence (AI) algorithms in applications impacting human users and actors has, as a direct consequence, the need for endowing these AI systems with ethical behaviors. While several approaches already exist, the question of adaptability to changes in contexts, users behaviors or preferences still remains open. We propose to tackle this question using Multi-Agent Reinforcement Learning of ethical behavior in different situations using Q-Tables and Dynamic Self-Organizing Maps to allow dynamic learning of the representation of the environment's state and reward functions to prescribe ethical behaviors. To evaluate this proposal, we developed a simulator of intelligent management of energy distribution in Smart Grids, evaluating different rewards functions to trigger ethical behaviors. Results show the ability to adapt to different conditions. Besides contributions on ethical adaptation, we compare our model to other learning approaches and show it performs better than a Deep Learning one (based on Actor-Critic). 

Citation:

Rémy Chaput, Olivier Boissier, Mathieu Guillermin, Salima Hassas. Apprentissage adaptatif de comportements éthiques. 28e Journées Francophones sur les Systèmes Multi-Agents (JFSMA'2020), Jun 2020, Angers, France.

BibTeX: 

@inproceedings{chaput:hal-03012127,
  TITLE = {{Apprentissage adaptatif de comportements {\'e}thiques}},
  AUTHOR = {Chaput, R{\'e}my and Boissier, Olivier and Guillermin, Mathieu and Hassas, Salima},
  URL = {https://hal.science/hal-03012127},
  BOOKTITLE = {{28e Journ{\'e}es Francophones sur les Syst{\`e}mes Multi-Agents (JFSMA'2020)}},
  ADDRESS = {Angers, France},
  PUBLISHER = {{C{\'e}padu{\`e}s}},
  YEAR = {2020},
  MONTH = Jun,
  KEYWORDS = {Ethics ; Reinforcement Learning ; Multi-Agent Systems ; Energy management ; {\'E}thique ; Apprentissage par renforcement ; Syst{\`e}mes Multi-Agent ; R{\'e}partition de l'{\'e}nergie},
  PDF = {https://hal.science/hal-03012127/file/JFSMA_2020_paper_12.pdf},
  HAL_ID = {hal-03012127},
  HAL_VERSION = {v1},
}