22. 22
2. Background
背景のまとめ
著者は前論文[1]でCSSを効率よく求める
Optimistic Linear Support(OLS)という
アルゴリズムを提案した
が、スカラー化したMDPをいちいち解くのは時間がかかる
[1] Roijers, D. M, et al. (2015). Computing convex coverage sets for faster multi-objective
coordination. Journal of Artificial Intelligence Research, 52, 399-443.
23. 23
2. Background
背景のまとめ
DQNならネットワークの重みを再利用できて
学習が効率的なのでは?
…というのが今回の提案
Deep Optimistic Linear Support Learning(DOL)
[1] Roijers, D. M, et al. (2015). Computing convex coverage sets for faster multi-objective
coordination. Journal of Artificial Intelligence Research, 52, 399-443.
著者は前論文[1]でCSSを効率よく求める
Optimistic Linear Support(OLS)という
アルゴリズムを提案した
が、スカラー化したMDPをいちいち解くのは時間がかかる
43. 43
Appendix
Much existing research assumes the Pareto coverage set(PCS), or Pareto front,
as the optimal solution set. However, we argue that this is not always the best choice.
… Because CCSs are typically much smaller, and have exploitable mathematical
properties, CCSs are often much cheaper to compute than PCSs.
“Efficient Methods for Multi-Objective Decision-Theoretic Planning”
Diederik M. Roijers. IJCAI. 2015.
In practice, the PCS and the CCS are often equal to the PF and CH. However,
the algorithms proposed in this article are guaranteed to produce a PCS or a CCS,
and not necessarily the entire PF or the CH. Because PCSs and the CCSs are
sucient solutions in terms of scalarized value, we say that these algorithms solve the
MO-CoGs.
"Computing Convex Coverage Sets for Faster Multi-objective Coordination.".
Diederik M. Roijers. Intell. Res.(JAIR) 52. 2015.
Pareto Coverage Setを求めないことについて