2020/04/16 by János Kramár, Kramár, János, Neil C. Rabinowitz +5 · 1 citation
Computer Science · Physics and Astronomy · Social Sciences · #FOS: Computer and information sciences #Human Mobility and Location-Based Analysis #Multiagent Systems (cs.MA) #Opinion Dynamics and Social Influence #Reinforcement Learning in Robotics
paper · pdf · doi:10.48550/arxiv.2004.07625
openalex publication_date 2020/04/16 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
One of the fundamental challenges of governance is deciding when and how to\nintervene in multi-agent systems in order to impact group-wide metrics of\nsuccess. This is particularly challenging when proposed interventions are novel\nand expensive. For example, one may wish to modify a building's layout to\nimprove the efficiency of its escape route. Evaluating such interventions would\ngenerally require access to an elaborate simulator, which must be constructed\nad-hoc for each environment, and can be prohibitively costly or inaccurate.\nHere we examine a simple alternative: Optimize By Observational Extrapolation\n(OBOE). The idea is to use observed behavioural trajectories, without any\ninterventions, to learn predictive models mapping environment states to\nindividual agent outcomes, and then use these to evaluate and select changes.\nWe evaluate OBOE in socially complex gridworld environments and consider novel\nphysical interventions that our models were not trained on. We show that neural\nnetwork models trained to predict agent returns on baseline environments are\neffective at selecting among the interventions. Thus, OBOE can provide guidance\nfor challenging questions like: "which wall should I tear down in order to\nminimize the Gini index of this group?"\n