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

定規とコンパスと ChainerRL

horiem
June 09, 2018

定規とコンパスと ChainerRL

強化学習を使って作図問題を解く
Chainer Meetup #07, 9th Jun 2018

horiem

June 09, 2018
Tweet

More Decks by horiem

Other Decks in Technology

Transcript

  1. ఆنͱίϯύεͱ
    ChainerRL
    Chainer Meetup #07, 9th Jun 2018
    [email protected]

    View Slide

  2. ChainerRL Ͱ

    View Slide

  3. ࡞ਤ໰୊Λղ͖͍ͨ

    View Slide

  4. ࡞ਤ໰୊
    • ఆنͱίϯύε͚ͩΛ࢖ͬͯ໨తͷਤܗΛඳ͘໰୊
    http://mathworld.wolfram.com/GeometricConstruction.html

    View Slide

  5. σϞ

    View Slide

  6. ͷલʹ

    View Slide

  7. ਤͷݟํ
    ֶशϞσϧʹ౉͢৘ใ
    ʢObservationʣ
    ਓؒ༻
    ໨తͷਤܗ
    ར༻Մೳͳ

    View Slide

  8. σϞ

    View Slide

  9. શମ૾
    ؀ڥ
    ΤʔδΣϯτ
    ߦಈ
    ؍ଌ

    View Slide

  10. શମ૾
    ؀ڥ
    ΤʔδΣϯτ
    [p0_x, p0_y]
    [p1_x, p1_y]
    ……
    ը૾
    ఺ͷ৘ใ
    [shape_flag, pi, pj]
    ৽͍͠ਤܗ

    View Slide

  11. ωοτϫʔΫΞʔΩςΫνϟ
    Conv
    MLP
    MLP
    Conv
    MLP
    [p0_x, p0_y]
    [p1_x, p1_y]
    ……
    ը૾
    ఺ͷ৘ใ
    [shape_flag, pi, pj]
    ৽͍͠ਤܗ

    View Slide

  12. ωοτϫʔΫΞʔΩςΫνϟ
    (100, 100)
    (12, 3)
    Conv
    MLP
    MLP
    Conv
    MLP
    (2, 12, 12)
    [p0_x, p0_y]
    [p1_x, p1_y]
    ……
    ը૾
    ఺ͷ৘ใ
    [shape_flag, pi, pj]
    ৽͍͠ਤܗ

    View Slide

  13. ωοτϫʔΫΞʔΩςΫνϟ
    (100, 100)
    (12, 3)
    Conv
    MLP
    MLP
    Conv
    MLP
    (2, 12, 12)
    [p0_x, p0_y]
    [p1_x, p1_y]
    ……
    ը૾
    ఺ͷ৘ใ
    [shape_flag, pi, pj]
    ৽͍͠ਤܗ
    = 288

    View Slide

  14. ࢥͬͨ͜ͱͳͲ
    • ڧԽֶश΍ͬͨ͜ͱͳ͔͚ͬͨͲָ͍͠
    • ChainerRL ϥΫͰΑ͍
    • ߦಈۭ͕ؒେ͖͍ͷͰݮΒ͍ͨ͠
    • AlphaGO ͕ࢀߟʹͳΔ͔΋ʁ
    • ίʔυ͖Ε͍ʹͨ͠Βެ։ && ղઆ͠·͢
    • n ࣍ํఔࣜΛ ChainerRL Ͱղ͚Δ͔ʁ
    • ՝֎׆ಈ޷͖ͳਓɺҰॹʹ΍Γ·͠ΐ͏ʂ

    View Slide