Using Reinforcement Learning to Learn Novel Strategies for Collective Decision Making

Abstract In military settings, collaborative decision making is often used to solve large and complex problems under time pres- sure. Here, workers can adopt particularly promising solutions found by colleagues (imitation) or independently explore new solutions (innovation). However, there exists a trade-off between imitation and individual innovation which has a consequential impact on the quality of the final solution found. Therefore, the design of effective collaboration strategies is an important problem when trying to find good solutions to large and complex problems. This paper formulates this strategy design as a reinforcement learning problem and presents preliminary results showing that, over short time periods, reinforcement learning outperforms most handcrafted heuristics that are typically used in these settings.
  • Hugo McNally (Southampton)
  • Sebastian Stein (Southampton)
  • Malgorzata Turalska (ARL)
  • Rosie Lickorish (IBM UK)
  • Geeth De Mel (IBM UK)
Date Sep-2020
Venue 4th Annual Fall Meeting of the DAIS ITA, 2020