Introduction to Domain Reduction


Consider the following problem: \begin{equation} \begin{aligned} P:= \min_{x_1,x_2} & f(x_1,x_2) \\ \text{s.t. } & x_1x_2 \leq 1,\\ & -10x_1-x_2 \leq -6,\\ & x_1 \in [0,1.5],\\ & x_2 \in [1,4]. \end{aligned} \end{equation} The feasible region for this problem corresponds to the red area in Figure 1 (denoted as \(\Omega\)). Notice that although the range of the variable \(x_1\) is originally \([0,1.5]\), the feasible region in the figure only admits values of \(x_1\) between \(0.2\) and \(1\) (red dashed lines in Figure 1). Knowing this, it is then logical to change the bounds of \(x\) in the original problem from \(x_1 \in [0,1.5]\) to \(x_1 \in [0.2,1]\) (i.e., tightening the bounds of \(x_1\)). This change will make other calculations and techniques used by Octeract Engine more effective (e.g., relaxations and branching).

Figure 1: domain reduction
Figure 1.

In general, for a problem \(\min_{x \in \Omega} f(x)\) we would like to find the tightest bounds for each variable \(x_i\), i.e., find the lower bound \(l_i^\star = \min_{x \in \Omega} x_i\), and upper bound \(u_i^\star = \max_{x \in \Omega} x_i\).

However, unlike our previous example, it is not easy to solve the problems \(\min_{x \in \Omega} x_i\) and \(\max_{x \in \Omega} x_i\), and a different approach is required. Feasibility based bounds tightening (FBBT) , Optimisation based bounds tightening (OBBT) and Constraint propagation (CP), are the algorithmic techniques used by Octeract Engine to approach this important task.
dr/dr1000