sisl/NeuralVerification.jl

Local Search in Sherlock can be unbounded

changliuliu opened this issue · 2 comments

Sherlock takes in HPolytope input, which can be unbounded. The local search function can be unbounded. For example, if we have a function y = x with constraint x>=0, the maximum on y is infinity. Global search should be fine since it only solves a feasibility problem.

  • One easy fix is to require the input set to be bounded.
  • A better fix is to explicitly check for unbounded cases.

Let's check what happens if the optimization in local_search is unbounded. Is neurons[1] in that case Inf/NaN/something else? And also is this easily detectible with INFEASIBLE_OR_UNBOUNDED. We can use the y = x test case. For now, we can make sure Sherlock is bounded along with Planet and Duality.

Closed by #151