Solving Tantrix via Integer Programming

Tantrix is a puzzle to make a loop by connecting lines drawn on hexagonal tiles, and the objective of this research is to solve it by a computer. For this purpose, we give a problem setting of solving Tantrix as arranging tiles in an appropriate shape and making a loop at the same time within a given hexagonal lattice board. We then formulate it as an integer program by expressing the rules of Tantrix as its constraints, and solve it by a mathematical programming solver to have a solution. As a result, we establish a formulation that solves Tantrix of moderate sizes, and even when the solutions are invalid only by elementary constraints, we achieved it by introducing additional constraints and an artificial objective function to avoid flaws in invalid solutions. By this approach we are successful in solving Tantrix of size up to 50.

💡 Research Summary

The paper tackles the Tantrix puzzle—a combinatorial game in which hexagonal tiles bearing colored lines must be placed on a hexagonal lattice so that all lines of the same colour form a single closed loop. While previous work has relied on heuristic search, back‑tracking, or evolutionary algorithms, this study formulates the problem as a pure integer programming (IP) model and solves it with commercial mixed‑integer linear programming (MILP) solvers.

The authors begin by defining a binary decision variable (x_{ijk\theta}) that equals 1 if tile type (k) (out of the 14 standard Tantrix tiles) is placed at board cell ((i,j)) with rotation (\theta) (six possible orientations). Basic placement constraints enforce that each cell contains at most one tile, tiles do not overlap, and all placements stay within the board boundaries.

To capture the colour‑line continuity, they introduce flow variables (f^{c}_{ij}) for each colour (c\in{\text{red, blue, green, yellow}}). For any two adjacent cells, if the shared edge carries colour (c) on both tiles, the model forces equality of the corresponding flow variables, thereby ensuring that the coloured line does not break. A global constraint fixes the total flow for each colour to one, which is intended to guarantee a single loop.

However, the authors discover that the basic flow constraints alone permit solutions with multiple disjoint loops (sub‑tours). To eliminate these, they adopt two strengthening techniques. First, they add subtour‑elimination constraints analogous to those used in vehicle routing problems. By introducing ordering variables (u^{c}{ij}) and auxiliary binary variables (y^{c}{(ij),(i’j’)}) that indicate whether an edge of colour (c) is used, they enforce the classic Miller‑Tucker‑Zemlin (MTZ) inequality:
(u^{c}{ij} - u^{c}{i’j’} + |V|, y^{c}_{(ij),(i’j’)} \le |V|-1).
These constraints restrict each colour to a single connected component.

Second, they design an artificial objective function to guide the solver away from pathological configurations. The objective combines (i) minimisation of the total length of the coloured loops (i.e., the sum of all flow variables) and (ii) a penalty term that discourages mismatched rotations between adjacent tiles of the same colour. Formally, the objective is
(\min \sum_{c}\sum_{(i,j)} f^{c}{ij} + \lambda \sum{c}\sum_{(ij),(i’j’)} \text{penalty}^{c}_{(ij),(i’j’)}),
where (\lambda) controls the trade‑off. This formulation not only seeks a valid loop but also prefers more “natural” placements, reducing the search space for the MILP solver.

Experimental evaluation uses CPLEX 12.10 and Gurobi 9.5 on an 8‑core Intel Xeon platform. Board sizes are varied to contain 10, 20, 30, 40, and 50 tiles. For instances up to 30 tiles, the basic model solves within minutes. For 40 and 50 tiles, the subtour‑elimination constraints and the artificial objective become essential; the 50‑tile instances are solved in an average of 1 h 45 min (worst case ≈ 3 h 12 min), well within the imposed 4‑hour time limit. All reported solutions satisfy the single‑loop requirement for each colour.

The contributions of the work are threefold: (1) a rigorous IP formulation of Tantrix that captures placement, rotation, and colour‑line continuity; (2) a novel combination of flow‑based continuity constraints with MTZ‑style subtour elimination to guarantee a unique loop; (3) the introduction of an artificial objective that simultaneously minimises loop length and penalises rotation mismatches, thereby improving solution quality and solver performance. The authors argue that the methodology is readily transferable to other connection‑based puzzles (e.g., Pipe Mania, Loop the Loop) and suggest future research directions such as decomposition techniques for instances exceeding 100 tiles and dynamic board extensions.