Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Safe Reinforcement Learning - Guest Speaker Ses...

Akifumi Wachi
February 13, 2024

Safe Reinforcement Learning - Guest Speaker Session at Cohere for AI

This is the slide deck in the Guest Speaker Session at Cohere for AI.
- https://cohere.com/events/c4ai-Akifumi-Wachi-2023

Akifumi Wachi

February 13, 2024
Tweet

More Decks by Akifumi Wachi

Other Decks in Research

Transcript

  1. "HFOEB • 4BGF3FJOGPSDFNFOU-FBSOJOHJOB/VUTIFMM • 8BDIJ l"4VSWFZPG$POTUSBJOU'PSNVMBUJPOTJO4BGF3FJOGPSDFNFOU-FBSOJOHz • http://arxiv.org/abs/2402.02025 (Released on

    Feb 6, 2024) • 1SFTFOUUXPQBQFST • 8BDIJBOE4VJ l4BGF3FJOGPSDFNFOU-FBSOJOHJO$POTUSBJOFE.BSLPW%FDJTJPO 1SPDFTTFTz *O*$.- • IUUQTQSPDFFEJOHTNMSQSFTTWXBDIJB • 8BDIJ l4BGF&YQMPSBUJPOJO3FJOGPSDFNFOU-FBSOJOH"(FOFSBMJ[FE'PSNVMBUJPO BOE"MHPSJUINTz *O/FVS*14 • IUUQTPQFOSFWJFXOFUGPSVN JEE2-TW,/X;$ 
  2. "HFOEB • 4BGF3FJOGPSDFNFOU-FBSOJOHJOB/VUTIFMM • 8BDIJ l"4VSWFZPG$POTUSBJOU'PSNVMBUJPOTJO4BGF3FJOGPSDFNFOU-FBSOJOHz • http://arxiv.org/abs/2402.02025 (Released on

    Feb 6, 2024) • 1SFTFOUUXPQBQFST • 8BDIJBOE4VJ l4BGF3FJOGPSDFNFOU-FBSOJOHJO$POTUSBJOFE.BSLPW%FDJTJPO 1SPDFTTFTz *O*$.- • IUUQTQSPDFFEJOHTNMSQSFTTWXBDIJB • 8BDIJ l4BGF&YQMPSBUJPOJO3FJOGPSDFNFOU-FBSOJOH"(FOFSBMJ[FE'PSNVMBUJPO BOE"MHPSJUINTz *O/FVS*14 • IUUQTPQFOSFWJFXOFUGPSVN JEE2-TW,/X;$ 
  3. 4BGFUZ*TTVFTJO3-  (BQ 3FTFBSDIPO3- 3FBM"QQMJDBUJPOT "MQIB(P (PPHMF%FFQ.JOE "VUPOPNPVT%SJWJOH (SBO5VSJTNP 4POZ"*

    6OTBGFBDUJPOTˠ/P1SPCMFNJ 6OTBGFBDUJPOTˠ$BUBTUSPQIJDSFTVMUTL .FEJDBM "QQMJDBUJPOT 4BGF3-JTOFFEFE
  4. 4BGF3-JO5IJT5BML • 4BGF3-JTBCSPBEUPQJDCZEFGJOJUJPO • (BSDJBBOE'FSOÂOEF[  DMBTTJGJFEPQUJNJ[BUJPODSJUFSJBJOUPGPVSHSPVQT  $POTUSBJOFEDSJUFSJPO 

    8PSTUDBTFDSJUFSJPO  3JTLTFOTJUJWFDSJUFSJPO  0UIFST FH STRVBSFE WBMVFBUSJTL • 5IJTUBMLGPDVTFTPOTBGF3-CBTFEPOUIFDPOTUSBJOFEDSJUFSJPO  Garcıa and Fernández. "A comprehensive survey on safe reinforcement learning." JMLR 16.1 (2015): 1437-1480.
  5. $PNNPO$POTUSBJOU3FQSFTFOUBUJPOT  TVCKFDUUP 4BGFUZ$POTUSBJOU 5ZQJDBM3-PCKFDUJWF 1SPCMFN'PSNVMBUJPO &YQFDUFE$VNVMBUJWF4BGFUZ$POTUSBJOU • 0OFPGUIFNPTUQPQVMBSGPSNVMBUJPOT •

    .BOZXFMMLOPXOBMHPSJUINTBSFCBTFEPOUIJTGPSNVMBUJPO • $10<> \5310 110^-BHSBOHJBO<> 3$10<> FUD • 'PDVTPOUIFBWFSBHFEQFSGPSNBODFˠ3FMBUJWFMZMPXSFRVJSFETBGFUZMFWFM <>"DIJBN $POTUSBJOFEQPMJDZPQUJNJ[BUJPO*O*$.-  <>3BZ #FODINBSLJOHTBGFFYQMPSBUJPOJOEFFQSFJOGPSDFNFOUMFBSOJOHBS9JWQSFQSJOUBS9JW  <>5FTTMFS 3FXBSEDPOTUSBJOFEQPMJDZPQUJNJ[BUJPO*O*$-3 
  6. $PNNPO$POTUSBJOU3FQSFTFOUBUJPOT  1SPCMFN'PSNVMBUJPO "MNPTU4VSFMZ$VNVMBUJWF4BGFUZ$POTUSBJOU TVCKFDUUP 4BGFUZ$POTUSBJOU 5ZQJDBM3-PCKFDUJWF • 3FRVJSFUIFDPOTUSBJOUTBUJTGBDUJPOXJUIQSPCBCJMJUZPG JF

    BMNPTUTVSFMZ  • ℙ! JTVTFESBUIFSUIBO𝔼!  • )JHIFSSFRVJSFEMFWFMPGTBGFUZ • 4BVUF3-<>BMHPSJUINJTCBTFEPOUIJTGPSNVMBUJPO • (PPEUIFPSFUJDBMQSPQFSUJFT &NQJSJDBMQFSGPSNBODF <>4PPUMB 4BVUF3-"MNPTUTVSFMZTBGFSFJOGPSDFNFOUMFBSOJOHVTJOHTUBUFBVHNFOUBUJPO*O*$.- 
  7. $PNNPO$POTUSBJOU3FQSFTFOUBUJPOT  1SPCMFN'PSNVMBUJPO "MNPTU4VSFMZ*OTUBOUBOFPVT 4BGFUZ$POTUSBJOU TVCKFDUUP 4BGFUZ$POTUSBJOU 5ZQJDBM3-PCKFDUJWF • 3FRVJSFUIFDPOTUSBJOUTBUJTGBDUJPOXJUIQSPCBCJMJUZPGBUFWFSZUJNFTUFQ

    • 7FSZIJHISFRVJSFEMFWFMPGTBGFUZ • .BOZBMHPSJUINTBSFCBTFEPOUIJTGPSNVMBUJPO • 4.C3-<> 3-$#'<> 4BGF.%1<> 4/0.%1<> FUD <>#FSLFOLBNQ 4BGFNPEFMCBTFESFJOGPSDFNFOUMFBSOJOHXJUITUBCJMJUZHVBSBOUFFT*O/FVS*14  <>$IFOH &OEUPFOETBGFSFJOGPSDFNFOUMFBSOJOHUISPVHICBSSJFSGVODUJPOTGPSTBGFUZDSJUJDBMDPOUJOVPVTDPOUSPMUBTLT*O"""*  <>5VSDIFUUB 4BGFFYQMPSBUJPOJOGJOJUF.BSLPWEFDJTJPOQSPDFTTFTXJUI(BVTTJBOQSPDFTTFT*O/FVS*14  <>8BDIJBOE4VJ4BGFSFJOGPSDFNFOUMFBSOJOHJODPOTUSBJOFE.BSLPWEFDJTJPOQSPDFTTFT*O*$.- 
  8. $PNNPO$POTUSBJOU3FQSFTFOUBUJPOT  • .BOZPUIFSEJWFSTFDPOTUSBJOUSFQSFTFOUBUJPOT FYDFQUGPSUIFUISFF • +PJOUDIBODFDPOTUSBJOU 4UBUFDPOTUSBJOU FUD •

    $POTUSBJOUGPSNVMBUJPO JTUIFGJSTUTUFQJOTBGF3- • $SVDJBMUPQSPQFSMZVOEFSTUBOEEJWFSTFDPOTUSBJOUSFQSFTFOUBUJPOT • 8F`WFSFDFOUMZSFMFBTFEBTVSWFZQBQFSPOBS9JWBCPVUDPOTUSBJOUSFQSFTFOUBUJPOT • 8BDIJ 4IFO BOE4VJ"4VSWFZPG$POTUSBJOU'PSNVMBUJPOTJO4BGF3FJOGPSDFNFOU -FBSOJOHzBS9JWQSFQSJOUBS9JW   1SPCMFN'PSNVMBUJPO )PXUPEFGJOFUIFTBGFUZDPOTUSBJOU 1PMJDZ0QUJNJ[BUJPO 6TFBQSPQFSBMHPSJUIN
  9. "HFOEB • 4BGF3FJOGPSDFNFOU-FBSOJOHJOB/VUTIFMM • 8BDIJ l"4VSWFZPG$POTUSBJOU'PSNVMBUJPOTJO4BGF3FJOGPSDFNFOU-FBSOJOHz • http://arxiv.org/abs/2402.02025 (Released on

    Feb 6, 2024) • 1SFTFOUUXPQBQFST • 8BDIJBOE4VJ l4BGF3FJOGPSDFNFOU-FBSOJOHJO$POTUSBJOFE.BSLPW%FDJTJPO 1SPDFTTFTz *O*$.- • IUUQTQSPDFFEJOHTNMSQSFTTWXBDIJB • 8BDIJ l4BGF&YQMPSBUJPOJO3FJOGPSDFNFOU-FBSOJOH"(FOFSBMJ[FE'PSNVMBUJPO BOE"MHPSJUINTz *O/FVS*14 • IUUQTPQFOSFWJFXOFUGPSVN JEE2-TW,/X;$ 
  10. 1SPCMFN4UBUFNFOU  • 3FXBSEGVODUJPO𝑟BOETBGFUZGVODUJPO𝑔 BSF VOLOPXOBQSJPSJ • *UJTJOUSBDUBCMFUPTPMWFUIJTQSPCMFN XJUIPVUGVSUIFSBTTVNQUJPOT •

    5XPBTTVNQUJPOTGSPN4VJFUBM  BOE5VSDIFUUBFUBM   • "TTVNQUJPO"HFOUTUBSUTJOBTFUPGTBGFTUBUFT XIJDIJTLOPXOUPCFTBGF • "TTVNQUJPO3FXBSEBOETBGFUZGVODUJPOTFYIJCJUSFHVMBSJUZ Ø 8FNPEFMUIFNVTJOH(BVTTJBO1SPDFTTFT 4VJFUBM "4BGFFYQMPSBUJPOGPSPQUJNJ[BUJPOXJUI(BVTTJBOQSPDFTTFT*O*$.-  5VSDIFUUBFUBM "4BGFFYQMPSBUJPOJOGJOJUF.BSLPWEFDJTJPOQSPDFTTFTXJUI(BVTTJBOQSPDFTTFT*O/FVS*14 
  11. .BJO$POUSJCVUJPO  8BDIJFUBMl4BGF&YQMPSBUJPOBOE0QUJNJ[BUJPOPG $POTUSBJOFE.%1TVTJOH(BVTTJBO1SPDFTTFTz """* • 4BGFUZQSPCBCJMJTUJDHVBSBOUFF • 0QUJNBMJUZOPHVBSBOUFF 5IJTQBQFS4BGF/FBSPQUJNBM.%1

    4/0.%1 8BDIJBOE4VJ l4BGF3FJOGPSDFNFOU-FBSOJOH JO$POTUSBJOFE.BSLPW%FDJTJPO1SPDFTTFTz*$.- • 4BGFUZQSPCBCJMJTUJDHVBSBOUFF • 0QUJNBMJUZQSPCBCJMJTUJDHVBSBOUFF 8BDIJ 4BGFFYQMPSBUJPOBOEPQUJNJ[BUJPOPGDPOTUSBJOFE.%1TVTJOHHBVTTJBOQSPDFTTFT *O"""*
  12. 4UFQXJTF"QQSPBDI  5IJTXPSLUBLFTBTUFQXJTFBQQSPBDI  &YQMPSBUJPOPGTBGFUZ  0QUJNJ[BUJPOPGUIFDVNVMBUJWFSFXBSE JOUIFDFSUJGJFETBGFSFHJPO &YQMPSBUJPO PG4BGFUZ

    &YQMPJUBUJPO PG3FXBSE &YQMPSBUJPO PG3FXBSE 4VQQPTFBOBHFOUDBOTVGGJDJFOUMZFYQBOEUIFTBGFSFHJPO 5IFO UIFBHFOUPOMZIBTUPPQUJNJ[FUIFDVNVMBUJWFSFXBSEJOUIFDFSUJGJFETBGFSFHJPO 8BDIJFUBM  USJFEUPTPMWFUIF UISFFXBZUSBEFPGGTJNVMUBOFPVTMZ 8BDIJ 4BGFFYQMPSBUJPOBOEPQUJNJ[BUJPOPGDPOTUSBJOFE.%1TVTJOHHBVTTJBOQSPDFTTFT *O"""*
  13. 4UFQ&YQMPSBUJPOPG4BGFUZ  ʔTBGFUZGVODUJPO ʔUISFTIPME ɾPCTFSWBUJPO  &YQBOTJPOPG4BGF3FHJPO • 5IFFGGJDJFODZPGFYQBOEJOHUIFTBGFSFHJPOJTNFBTVSFE CZUIFXJEUIPGUIFTBGFUZGVODUJPO`TDPOGJEFODFJOUFSWBM

    • 4BNQMFUIFOFYUTUBUFXJUIUIFNBYJNVN XJUIJOUIFTBGFTQBDF  $POWFOUJPOBM5FSNJOBM$POEJUJPOPG&YQMPSBUJPOPG4BGFUZ • 5IFQSFWJPVTXPSL 4VJFUBM 5VSDIFUUBFUBM  UFSNJOBUFEUIF FYQMPSBUJPOJGUIFGPMMPXJOHFRVBUJPOIPMETGPSBMMTUBUFTJOTBGFTQBDF
  14. &4"MHPSJUIN  • $POTJEFSBOFX.%1 XIFSFSFXBSEGVODUJPOJTEFGJOFEBTJOUIFGJHVSFCFMPX • 3FXBSEJTTFUUPCFUIFMPXFSCPVOEJOUIFDVSSFOUMZJEFOUJGJFETBGFSFHJPO • 0UIFSXJTF TFUUPCFUIFVQQFSCPVOE

    • 5IJTSFXBSETFUUJOHTFODPVSBHFUIFBHFOUUPFYQMPSFPVUTJEFTUIFDVSSFOUMZ JEFOUJGJFETBGFSFHJPO 6ODFSUBJO SFHJPO *EFOUJGJFETBGF SFHJPO 3FXBSEMPXFSCPVOE 3FXBSE VQQFSCPVOE *EFOUJGJFE VOTBGFSFHJPO • 4VQQPTFUIFTFUPGOFYUTUBUFTUIBU UIFBHFOUXJMMWJTJUCBTFEPOUIF PQUJNBMQPMJDZJTBTVCTFUPGUIF DVSSFOUMZJEFOUJGJFETBGFSFHJPO • 4UPQFYQMPSJOHUIFTBGFUZGVODUJPO
  15. &YQFSJNFOUT  • %FWFMPQFEBOFXTJNVMBUJPOFOWJSPONFOU DBMMFE(14BGFUZ(ZN XIJDI JTCBTFEPO0QFO"*4BGFUZ(ZN 3BZ FU BM

      • "DIJFWFECFUUFSFNQJSJDBMQFSGPSNBODFUIBOPUIFSCBTFMJOFT • 4BGF.%1 5VSDIFUUBFUBM  • 4BGF&YQ0QU.%1 8BDIJFUBM  3FXBSE IJHI ZFMMPX MPX CMVF 4BGFUZIFJHIU
  16. &GGFDUPG&4"MHPSJUIN  • 1VSFTUFQXJTFBQQSPBDI CSPXOMJOF IBTBOJTTVF • 4QFOENVDIUJNFGPSUIFFYQMPSBUJPOPGTBGFUZ • 8FBEEJUJPOBMMZQSPQPTFEFBSMZTUPQQJOHPGFYQMPSBUJPOPGTBGFUZ

    &4  • &4NBJOUBJOTUIFUIFPSFUJDBMHVBSBOUFFTPOOFBSPQUJNBMJUZ • 1&4FNQJSJDBMMZQFSGPSNTCFUUFSUIBO&4 &YQMPSBUJPOPG 4BGFUZ
  17. "HFOEB • 4BGF3FJOGPSDFNFOU-FBSOJOHJOB/VUTIFMM • 8BDIJ l"4VSWFZPG$POTUSBJOU'PSNVMBUJPOTJO4BGF3FJOGPSDFNFOU-FBSOJOHz • http://arxiv.org/abs/2402.02025 (Released on

    Feb 6, 2024) • 1SFTFOUUXPQBQFST • 8BDIJBOE4VJ l4BGF3FJOGPSDFNFOU-FBSOJOHJO$POTUSBJOFE.BSLPW%FDJTJPO 1SPDFTTFTz *O*$.- • IUUQTQSPDFFEJOHTNMSQSFTTWXBDIJB • 8BDIJ l4BGF&YQMPSBUJPOJO3FJOGPSDFNFOU-FBSOJOH"(FOFSBMJ[FE'PSNVMBUJPO BOE"MHPSJUINTz *O/FVS*14 • IUUQTPQFOSFWJFXOFUGPSVN JEE2-TW,/X;$ 
  18. ."4&"MHPSJUIN  $POTUSVDUBOVODFSUBJOUZRVBOUJGJFS FH (1 (-. $PNQVUFUIFTFUPG DPOTFSWBUJWFMZ TBGFBDUJPOT *G

    BHFOUUBLFTlFNFSHFODZTUPQBDUJPOzBOESFDFJWFTQFOBMUZ &YFDVUFUIFOFYUBDUJPOXJUIJOBOEUIFONPWFUP NBYJNJ[F TVCKFDUUP
  19. $PODMVTJPO 5BLF)PNF.FTTBHFT • 4BGFUZJTBOJNQPSUBOUJTTVFJO3- • %JGGFSFOUQSPCMFNTFUUJOHTˠ %JWFSTFDPOTUSBJOUSFQSFTFOUBUJPOT • 4JHOJGJDBOUUPVOEFSTUBOEJOUFSSFMBUJPOTCFUXFFOFBDIDPOTUSBJOUSFQSFTFOUBUJPO •

    1SFTFOUUXPTQFDJGJDBMHPSJUINT 4/0.%1 BOE."4&  • 4BGFUZ DBOCFHVBSBOUFFEXJUIIJHIQSPCBCJMJUZ CZDPOTUSVDUJOHBOVODFSUBJOUZRVBOUJGJFS WJB(1T (-.T • /FBSPQUJNBMJUZ DBOBMTPCFHVBSBOUFFEVOEFSQSPQFSBTTVNQUJPOT • ."4&JTDPNQBUJCMFXJUIBOZ VODPOTUSBJOFE EFFQ3-BMHPSJUIN • (PPEFNQJSJDBMQFSGPSNBODF  5IBOLZPV $POUBDUXBDIJBLJGVNJ<BU>HNBJMDPN