Idanwo ti o sopọ mọ itọsọna · Lile Ipele

Idanwo Ẹkọ Imudara Inverse

Checks the reward-recovery goal of IRL, ambiguity, maximum-entropy trajectory probabilities and how a policy is derived afterward.

Jẹmọ itọsọna awọn ọnaẸkọ Imudara onidakeji
Ibeere 1 ti 8

Kini ẹkọ imuduro onidakeji ṣe ifọkansi lati bọsipọ?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Kini ẹkọ imuduro onidakeji ṣe ifọkansi lati bọsipọ?
    • A ere iṣẹ ti o salaye šakiyesi iwé ihuwasi
    • A yiyara optimizer fun gradient ayalu
    • Ẹya fisinuirindigbindigbin ti awọn ifihan
    • Atokun aworan ti o ni aami
  2. Kini idi ti iṣoro IRL ṣe apejuwe bi aiṣedeede tabi aibikita?
    • Nitoripe awọn ifihan nigbagbogbo dara julọ
    • Nitori ọpọlọpọ awọn iṣẹ ere oriṣiriṣi le ṣe alaye ihuwasi kanna
    • Nitori awọn ere gbọdọ nigbagbogbo jẹ odi
    • Nitoripe ko nilo data amoye
  3. Anfani bọtini wo ni gbigbapada ere kan ni lori ti ẹda ihuwasi?
    • Ko nilo awọn ifihan
    • O nigbagbogbo nṣiṣẹ ni ibakan akoko
    • O gba ibi-afẹde ti o wa ni ipilẹ, gbogbogbo si awọn ipinlẹ ti a ko rii
    • O yago fun lilo eyikeyi eto imulo
  4. Bawo ni o pọju-entropy IRL yanju ambiguity ere?
    • Nipa ipa gbogbo awọn ere lati dogba
    • Nipa aibikita awọn itọpa suboptimal patapata
    • Nipa yiyọ iṣẹ ere
    • Nipa ṣiṣe apẹẹrẹ iṣeeṣe itọpa bi aropin ni ere lapapọ
  5. Lẹhin ti IRL gba iṣẹ ere kan pada, kini igbagbogbo ṣe atẹle?
    • Ẹkọ imuduro boṣewa n gba eto imulo lati ere yẹn
    • Ere ti wa ni asonu
    • Awọn ifihan ti paarẹ lati fi aaye pamọ
    • Aṣoju da kikọ ẹkọ duro
  6. Ninu oju iṣẹlẹ wo ni IRL wulo paapaa?
    • Nigbati ere naa ti mọ ni pipe
    • Nigbati o ba fẹ lati ni oye ati gbogbogbo awọn ibi-afẹde lẹhin ihuwasi iwé
    • Nigbati ko si data ihuwasi wa
    • Nigbati o nilo lati ṣe akori awọn iṣe ti o wa titi nikan
  7. Why can a constant zero reward explain many different expert policies?
    • It assigns the highest reward only to expert actions
    • It proves the expert is optimal
    • Every policy receives the same return, so none is preferred by the reward
    • It makes the transition model deterministic
  8. In maximum-entropy IRL, how are higher-reward trajectories treated relative to lower-reward ones?
    • They are assigned probability zero
    • They are treated as equally likely regardless of reward
    • They are removed before policy learning
    • They receive higher probability, with probability depending exponentially on total reward