{
  "id": "2104.13906",
  "version": "v1",
  "published": "2021-04-28T17:41:35.000Z",
  "updated": "2021-04-28T17:41:35.000Z",
  "title": "Reward (Mis)design for Autonomous Driving",
  "authors": [
    "W. Bradley Knox",
    "Alessandro Allievi",
    "Holger Banzhaf",
    "Felix Schmitt",
    "Peter Stone"
  ],
  "comment": "13 pages (25 pages with appendix), 4 figures",
  "categories": [
    "cs.LG"
  ],
  "abstract": "This paper considers the problem of reward design for autonomous driving (AD), with insights that are also applicable to the design of cost functions and performance metrics more generally. Herein we develop 8 simple sanity checks for identifying flaws in reward functions. The sanity checks are applied to reward functions from past work on reinforcement learning (RL) for autonomous driving, revealing near-universal flaws in reward design for AD that might also exist pervasively across reward design for other tasks. Lastly, we explore promising directions that may help future researchers design reward functions for AD.",
  "revisions": [
    {
      "version": "v1",
      "updated": "2021-04-28T17:41:35.000Z"
    }
  ],
  "analyses": {
    "subjects": [
      "91B16",
      "I.2.6",
      "I.2.8",
      "I.2.9"
    ],
    "keywords": [
      "autonomous driving",
      "reward design",
      "researchers design reward functions",
      "simple sanity checks",
      "past work"
    ],
    "note": {
      "typesetting": "TeX",
      "pages": 13,
      "language": "en",
      "license": "arXiv",
      "status": "editable"
    }
  }
}