escaping saddle points in constant dimensional spaces an
play

Escaping Saddle Points in Constant Dimensional Spaces: an - PowerPoint PPT Presentation

Escaping Saddle Points in Constant Dimensional Spaces: an Agent-based Modeling Perspective Grant Schoenebeck, University of Michigan Fang-Yi Yu , Harvard University Reinforced random walk with A discrete time stochastic process { :


  1. Escaping Saddle Points in Constant Dimensional Spaces: an Agent-based Modeling Perspective Grant Schoenebeck, University of Michigan Fang-Yi Yu , Harvard University

  2. Reinforced random walk with 𝐺 A discrete time stochastic process {π‘Œ 𝑙 : 𝑙 = 0, 1, … } in ℝ 𝑒 that admits the following representation, π‘Œ 𝑙 π‘Œ 𝑙+1 βˆ’ π‘Œ 𝑙 = 1 1 π‘œ 𝐺 π‘Œ 𝑙 + 𝑉 𝑙 π‘œ 𝐺(π‘Œ 𝑙 ) β€’ Agent based models with π‘œ agents 1 π‘œ 𝑉 𝑙 – Evolutionary games π‘Œ 𝑙+1 – Dynamics on social networks β€’ Heuristic local search algorithms with uniform step size 1/π‘œ

  3. Node Dynamic πŽπ„(𝐻, 𝑔 𝑢𝑬 , π‘Œ 𝟏 ) [SY18] β€’ Fixed a (weighted) graph 𝐻 = (π‘Š, 𝐹) 𝑠 π‘Œ π‘’βˆ’1 𝑀 = 1 7 opinion set {0,1} , an update function π’ˆ 𝑢𝑬 β€’ Given an initial configuration π‘Œ 0 :V ↦ {0,1} β€’ At round t, β€’ A node v is picked uniformly at random β€’ 𝒀 𝒖 π’˜ = 1 w.p. π’ˆ 𝑢𝑬 𝒔 𝒀 π’–βˆ’πŸ π’˜ ; = 0 otherwise

  4. Gradient-like dynamics Converges to an attracting fixed-point region in O(π‘œ log π‘œ) steps. If β€’ Noise, 𝑉 𝑙 – Martingale difference – bounded – Noisy β€’ Expected difference, 𝐺 ∈ π’Ÿ 2 – Fixed points are hyperbolic – Potential function

Recommend


More recommend