tf.contrib.constrained_optimization.find_best_candidate_distribution

View source on GitHub

Finds a distribution minimizing an objective subject to constraints.

tf.contrib.constrained_optimization.find_best_candidate_distribution(
    objective_vector, constraints_matrix, epsilon=0.0
)

This function deals with the constrained problem:

minimize f(w) s.t. g_i(w) <= 0 for all i in {0,1,...,m-1}

Here, f(w) is the "objective function", and g_i(w) is the ith (of m) "constraint function". Given a set of n "candidate solutions" {w_0,w1,...,w{n-1} }, this function finds a distribution over these n candidates that, in expectation, minimizes the objective while violating the constraints by the smallest possible amount (with the amount being found via bisection search).

The objective_vector parameter should be a numpy array with shape (n,), for which objective_vector[i] = f(w_i). Likewise, constraints_matrix should be a numpy array with shape (m,n), for which constraints_matrix[i,j] = g_i(w_j).

This function will return a distribution for which at most m+1 probabilities, and often fewer, are nonzero.

For more specifics, please refer to:

Cotter, Jiang and Sridharan. "Two-Player Games for Efficient Non-Convex Constrained Optimization". https://arxiv.org/abs/1804.06500

This function implements the approach described in Lemma 3.