---
title: Traveling Tournament Problem (TTP) - Wikitext Document
url: https://www.emergentmind.com/topics/traveling-tournament-problem-ttp
type: topic
---

# Traveling Tournament Problem (TTP) - Wikitext Document

The **Traveling Tournament Problem (TTP)** is a combinatorial optimization problem that combines double-round-robin sports scheduling with vehicle-routing decisions. Given an even number \(n\) of teams, each team must play every other team twice—once at each team’s home venue—over \(2(n-1)\) rounds, while minimizing the aggregate distance traveled by all teams. Teams begin and end at home and may travel directly between successive away venues. The standard formulation commonly imposes a no-repeat constraint and bounds consecutive home and away games by a parameter \(k\). The TTP originated in the scheduling challenges of Major League Baseball and was formally introduced by Easton, Nemhauser, and Trick in 2001 [2609.03612].

## 1. Formal model and objective

Let \(T=\{1,\ldots,n\}\) be an even-sized set of teams, each associated with a home venue. Let \(d_{ij}\) denote the distance between the venues of teams \(i\) and \(j\). The standard theoretical model assumes a symmetric metric:

\[
d_{ij}=d_{ji},\qquad d_{ii}=0,\qquad d_{ij}\leq d_{ik}+d_{kj}.
\]

A double round-robin tournament contains \(2(n-1)\) rounds. Each team plays exactly once in every round, and every unordered pair \(\{i,j\}\) meets twice: once with \(i\) at home and once with \(j\) at home. Consequently, the tournament contains \(n(n-1)\) games.

If team \(i\) visits venues \(v_{i,1},\ldots,v_{i,q}\) during an away sequence, its travel includes the journey from its home venue to \(v_{i,1}\), travel between consecutive away venues, and the return from \(v_{i,q}\) to its home venue. Summing over all road trips gives the total travel of team \(i\), and the TTP objective is

\[
\min \operatorname{Travel}(S)
 =\sum_{i=1}^{n}\operatorname{Travel}_i(S),
\]

where \(S\) is a complete schedule.

The standard constraints are:

- **Double round robin**: every ordered home-away pairing occurs exactly once.
- **One game per round**: every team plays exactly one game in each round.
- **No repeater**: the same two teams may not meet in consecutive rounds.
- **At-most-\(k\)**: no team may play more than \(k\) consecutive home games or more than \(k\) consecutive away games.

The classical benchmark is usually TTP(3), while TTP(2) imposes the more restrictive prohibition of \(HHH\) and \(AAA\). The unconstrained problem, commonly called the unconstrained TTP (UTTP), removes the no-repeater and home/away streak restrictions [1007.0501; 2609.03612].

The problem is difficult because a local scheduling decision simultaneously changes the home-away patterns of two teams, the road-trip decompositions of several teams, the no-repeat status of fixtures, and the feasibility of the global round-robin structure.

## 2. Variants and modeling assumptions

### Bounded-streak TTP

TTP-\(k\) limits every maximal home or away sequence to length \(k\). A road trip is therefore a consecutive sequence of at most \(k\) away games. The parameter changes both the combinatorial structure and the approximation behavior of the problem. TTP(2) has especially strong approximation results, whereas TTP(3) and larger fixed values are approximation-theoretically harder.

The TTP-2 case has been studied separately according to the parity of \(n/2\). For \(n/2\) even, equivalently \(n\equiv0\pmod 4\), a matching-based construction achieves an approximation ratio

\[
1+\frac{3}{n}-\frac{6}{n(n-2)}
\]

and runs in \(O(n^3)\) time [2108.13060]. A subsequent practical framework gives a \(1+3/n\) guarantee for even \(n/2\) and a \(1+5/n\) guarantee for odd \(n/2\), together with benchmark improvements [2212.12240]. For \(n\equiv2\pmod4\), a deterministic \(O(n^3)\)-time construction achieves a \(1+9/n\) ratio, improving earlier deterministic algorithms but not matching the randomized \(1+5/n\) result [2310.02592].

### Mirrored TTP

In the mirrored TTP (mTTP), the second half of the tournament repeats the opponent sequence of the first half with home and away reversed. If a first-half round contains \(i\) hosting \(j\), the corresponding second-half round contains \(j\) hosting \(i\). Mirroring restricts the schedule space and can simplify construction, although it is not part of the ordinary TTP definition [1311.1884].

Parallel simulated annealing has been applied to mTTP using independent OpenMP search trajectories. The reported throughput improvement is approximately \(1.5\)- to \(2.2\)-fold in solutions explored per unit time, with candidate solutions within \(10\%\) of known lower bounds on the tested instances [1311.1884]. Genetic algorithms have also been applied to mTTP by separating the home-away pattern from opponent scheduling. A swapping operation complements home-away sequences and preserves the mirrored structure. For team counts from \(4\) through \(20\), the resulting travel counts reached published lower bounds for \(n=4,6,8,10,12,18\), were one travel above the bound for \(n=14,20\), and six above it for \(n=16\) [1704.04879].

### Unconstrained TTP

The UTTP removes the no-repeater and at-most-\(k\) constraints. A constructive algorithm based on a Christofides Hamiltonian cycle, the circle method, systematic home-away assignment, and cyclic rotations provides a \(2.75\)-approximation for symmetric metric distances. If an optimal Hamiltonian cycle is available, the ratio improves to \(2.25\) [1110.0620].

The UTTP has also been studied polyhedrally. A cubic integer-programming formulation uses match variables \(x_{k,i,j}\) and travel variables \(y_{t,i,j}\). The resulting unconstrained traveling tournament polytope has dimension

\[
\dim P_{\mathrm{utt}(n)}=3n^3-8n^2+6n.
\]

The study identifies facet-defining lifted travel inequalities, distinct-venue flow inequalities, and home-flow inequalities. Computationally, the flow inequalities substantially strengthen linear-programming bounds, whereas some lifted inequalities have little practical effect despite being facet-defining [2011.09135].

### Linear and incomplete variants

The Linear Distance TTP (LD-TTP) assumes that venues lie on a line. If adjacent distances are \(d_1,\ldots,d_{n-1}\), total travel can be expressed through bridge-crossing coefficients:

\[
\operatorname{Cost}(S)=\sum_{k=1}^{n-1}c_k(S)d_k.
\]

For \(n=4\), the optimum is

\[
8(d_1+d_2+d_3),
\]

and for \(n=6\) the optimum is one of seven explicit expressions based on the bridge lengths. For \(n\equiv4\pmod6\), an expander construction gives a \(4/3\)-approximation for LD-TTP [1401.6049]. The same relaxation has been used heuristically for general metric instances.

The incomplete TTP (iTTP) reverses the temporal regime: there are too few rounds for every pair to meet, so the optimizer chooses both which games are played and when and where they occur. The iTTP is NP-hard even with only two rounds. Its formulations include independent and dependent lower bounds, road-trip-based integer programs, fixed-home-away-pattern models, and the 2-Team Home-AwaySwap neighborhood, which fully connects the feasible proper balanced home-away-pattern space [2603.19754].

## 3. Lower bounds and exact formulations

The most widely used lower bound is the independent lower bound (ILB), which optimizes each team’s routing problem independently and then sums the resulting values. For TTP(2), let

\[
D(i)=\sum_{j\neq i}d_{ij},\qquad
\Delta=\sum_{i=1}^{n}D(i),
\]

and let \(M\) be a minimum-weight perfect matching with weight \(d(M)\). An individual itinerary has lower bound \(D(i)+d(M)\), yielding

\[
\operatorname{OPT}\geq \Delta+n\,d(M).
\]

This bound is central to several approximation analyses [2108.08444; 2212.12240].

For UTTP, if \(\tau\) is the length of a shortest Hamiltonian cycle through all venues, then

\[
\operatorname{OPT}_{\mathrm{UTTP}}\geq n\tau.
\]

This bound underlies the \(2.75\)-approximation based on a Hamiltonian cycle [1110.0620].

The ILB ignores synchronization between teams. Consequently, independently optimal itineraries may select incompatible games, repeat the same pair, or fail to decompose into the required daily perfect matchings. Stronger bounds incorporate matching structures, spanning trees, aggregate distances, minimum travel legs, or joint road-trip compatibility.

For TTP(2), the independent lower bound

\[
LB=2D_G+nD_M
\]

uses \(D_G\), the total weight of the complete venue graph, and \(D_M\), the weight of a minimum perfect matching. For the iTTP, the dependent lower bound jointly selects road trips while enforcing that each pair meets at most once. A one-factorizable dependent lower bound additionally requires the selected games to be decomposable into rounds [2603.19754].

Exact methods include integer programming, branch-and-price, constraint programming, dynamic programming, and branch-and-bound. Branch-and-price and specialized search have established optimality for NL8 and NL10. Larger realistic instances remain difficult: NL12 has a reported best upper bound of 110,729 and lower bound of 108,629, leaving a gap of approximately \(2\%\) [2609.03612].

Direct game-based integer programs typically use binary variables indicating whether a game is scheduled in a round and auxiliary variables indicating travel between venues. The cubic formulation for UTTP is representative:

\[
x_{k,i,j}=1
\]

indicates that team \(j\) plays at venue \(i\) in slot \(k\), while

\[
y_{t,i,j}=1
\]

indicates that team \(t\) travels from venue \(i\) to venue \(j\). Consecutive-match constraints link \(x\) and \(y\), while first- and last-trip constraints model departure from and return to home.

Road-trip formulations use variables for complete ordered away trips. Their LP relaxations are stronger than direct travel-linking formulations, but the number of trip variables can grow as

\[
O(n^\lambda r),
\]

where \(\lambda\) is the maximum trip length and \(r\) is the number of rounds. This creates a trade-off between formulation strength and scalability [2603.19754].

## 4. Approximation algorithms

Approximation algorithms for the TTP exploit metric structure, matchings, Hamiltonian cycles, cycle packings, or carefully designed schedule blocks.

### Hamiltonian-cycle constructions

For UTTP, a Hamiltonian cycle obtained with Christofides’ algorithm is synchronized with a circle-method timetable and evaluated over all cyclic rotations. The method runs in \(O(n^3)\) time and achieves a \(2.75\)-approximation [1110.0620].

For general TTP-\(k\), a construction based on a \(3/2\)-approximate Hamiltonian cycle gives a \(5\)-approximation for every \(k\geq2\), improving to \(4\) when \(k\geq n/2\). The construction uses a random cyclic labeling, alternating directed blocks, and a boundary correction that prevents excessive streaks across the season boundary [2309.01902].

### Matching and super-team constructions

For TTP-2, minimum-weight perfect matchings identify pairs of teams that should ideally be visited during the same two-game road trip. These pairs are treated as super-teams. Super-games are expanded into four-day and six-day blocks whose home-away patterns avoid \(HHH\) and \(AAA\).

For \(n\equiv0\pmod4\), the main construction has a ratio

\[
1+\frac{3}{n}-\frac{6}{n(n-2)}
\]

and \(O(n^3)\) running time [2108.13060]. For \(n\equiv2\pmod4\), the construction requires special super-teams and right super-games, leading to a \(1+5/n\) randomized approximation [2212.12240]. The deterministic \(1+9/n\) construction selects exceptional matching pairs according to distance-sum values and avoids the \(O(n^4)\) derandomization cost [2310.02592].

A different construction for \(n\equiv4\pmod6\) uses two levels of matching and Type-1, Type-2, and Type-3 super-match blocks. Its stated guarantee is

\[
1+\frac{\left\lceil\log_2(n/4)\right\rceil+4}{2(n-2)}.
\]

The method controls the number of Type-2 “flip” blocks but does not optimize their distance contribution as effectively as their number [2109.09065].

### Cycle-packing constructions

Cycle packing replaces the single Hamiltonian-cycle viewpoint with a collection of \(k\)-cycles. Cycle-teams are grouped into super-teams, and normal, left, and right super-games are expanded through cyclic perfect matchings. The resulting schedules maintain no-repeat and bounded-streak conditions.

For TTP-3, combining cycle-packing schedules with Hamiltonian constructions yields a

\[
\frac{139}{87}+\varepsilon
\]

approximation. For TTP-4, the guarantee is

\[
\frac{17}{10}+\varepsilon.
\]

For constant \(k\geq5\), refined Hamiltonian analysis gives

\[
\frac{5k^2-4k+3}{2k(k+1)}+\varepsilon.
\]

For the linear-distance variant, cycle packing gives

\[
\frac{3k-3}{2k-1}+\varepsilon,
\]

which becomes \(6/5+\varepsilon\) for LDTTP-3 [2404.10955].

### Complexity barriers

TTP(2) admits a PTAS, although its exact complexity remains open. In contrast, TTP-\(k\) is APX-hard for every fixed \(k\geq3\), even with symmetric metric distances. The result is established through an L-reduction from a restricted \(k\)-Tour Cover problem. Consequently, unless \(\mathrm{P}=\mathrm{NP}\), no PTAS exists for fixed \(k\geq3\) [2308.14124].

For every fixed \(k>3\), TTP-\(k\) is strongly NP-complete. The reduction uses normalized \(k\)-SAT instances, variable and clause gadgets, colocated central teams, and a large metric scale that forces compact tours to encode satisfying assignments [2110.02300].

## 5. Heuristic and metaheuristic methods

Heuristic methods are important because exact optimization becomes difficult beyond approximately ten teams.

### Neighborhood search

A basic TTP schedule can be represented as a round-by-team matrix. Common neighborhood moves include:

- **SwapHomes**: reverse the home-away orientation of the two games between a pair.
- **SwapRounds**: exchange two complete rounds.
- **SwapTeams**: exchange the schedules of two teams.
- **PartialSwapRounds**: exchange selected games and propagate the changes through affected opponents.
- **PartialSwapTeams**: exchange a game and follow a compensating exchange chain until the double round-robin structure is restored.

The Lookahead Partial Swap Teams neighborhood extends PartialSwapTeams through

\[
swapHomes(t_j,t_l)
\rightarrow
partialSwapTeams(t_i,t_j,r)
\rightarrow
swapHomes(t_j,t_l).
\]

A swaplist records the successive opponents encountered by the exchange chain. Repeated opponents provide opportunities to shorten the chain. The resulting neighborhood can alter fewer games while reaching schedules inaccessible to ordinary partial swaps in one step. It was tested within simulated annealing and produced best-known values for Galaxy instances with 36, 38, and 40 teams, including improvements of \(14.5\%\), \(15.3\%\), and \(15.0\%\), respectively, over previous best-known values [1007.0501].

Simulated annealing generally accepts improving moves and probabilistically accepts worsening moves, with acceptance decreasing as temperature falls. Infeasible schedules may be allowed during the search, depending on the formulation and penalty mechanism.

### Parallel search

Independent simulated-annealing trajectories are naturally parallelizable. OpenMP implementations assign one trajectory to each thread and select the best thread-local solution after synchronization. This design has minimal communication overhead and increases the number of candidate schedules explored per unit time, although it does not permit threads to exchange discoveries during the search [1311.1884].

### Genetic algorithms

Genetic algorithms have been applied especially to mTTP by separating the home-away pattern from opponent assignment. A binary traveling table records home and away status, while a scheduling table records opponents. Complementary sequence generation ensures that paired teams have opposite home-away statuses and helps preserve feasibility during crossover. The swapping operation changes all home statuses to away and vice versa, preserving run lengths and the mirrored condition [1704.04879].

### Fairness-oriented objectives

The classical objective minimizes aggregate travel:

\[
\min \sum_i d_i.
\]

A Korean Baseball Organization application instead minimizes the standard deviation of team travel distances:

\[
\sigma_d=
\sqrt{\frac{1}{8}\sum_{i=1}^{8}(d_i-m)^2},
\qquad
m=\frac{1}{8}\sum_{i=1}^{8}d_i.
\]

Starting from the actual 2012 KBO schedule, simulated annealing with SwapRounds and SwapHomes reduced the reported travel-distance dispersion from approximately \(1{,}230\) km to \(142\) km. The optimized schedule also reduced total travel by approximately \(4.2\%\), although total travel was not the primary objective [1205.1299].

The fairness formulation demonstrates that TTP objectives need not be limited to aggregate transportation cost. However, a fairness objective may trade off against total travel, and simultaneous improvement is not guaranteed across instances.

## 6. Benchmarks, significance, and open problems

TTP research relies heavily on standardized benchmark families:

- **NL**: National League baseball locations.
- **NFL**: National Football League locations.
- **SUP**: Super Rugby locations.
- **BRA**: Brazilian soccer locations.
- **GAL**: three-dimensional galaxy coordinates.
- **CIRC**: teams arranged on a circle.
- **LINE**: equidistant teams on a line.
- **INCR**: teams on a line with increasing adjacent distances.
- **CON**: constant pairwise distances.

The RobinX sports-timetabling repository provides XML instances and schedules, feasibility validation, objective-value computation, and lower- and upper-bound tracking. Machine-checkable schedules are particularly important because published TTP results have sometimes omitted complete schedules, used inconsistent instance files, or reported insufficient computational details [2609.03612].

The major unresolved questions include:

- **Complexity of TTP(2)**: whether the problem is polynomial-time solvable or NP-hard.
- **Optimality beyond NL10**: whether NL12 or larger realistic instances can be solved to proven optimality.
- **Stronger lower bounds**: especially bounds that jointly represent routing, home-away patterns, and round decomposability.
- **Scalable exact methods**: including branch-and-price, column generation, Benders decomposition, and constraint programming.
- **Connected neighborhoods**: whether a polynomial-time neighborhood can connect the entire feasible schedule space while preserving double-round-robin feasibility.
- **Integrated objectives**: combining team travel, umpire travel, fairness, rest, and operational constraints.
- **Richer competition models**: including incomplete round robins, venue availability, broadcast requirements, multi-game series, and asymmetric or nonmetric travel.
- **Improved cycle and path packings**: because advances in minimum \(k\)-cycle or \(k\)-path packing directly improve TTP-\(k\) approximation ratios.

The TTP remains a canonical example of tightly coupled routing and scheduling. Its difficulty does not arise solely from the number of games or the size of the distance matrix. Rather, it results from the need to synchronize many individually constrained itineraries into a single sequence of daily perfect matchings. This structure has supported a broad research program spanning integer programming, polyhedral combinatorics, approximation algorithms, complexity theory, metaheuristics, parallel search, and sports-scheduling applications.

Source: https://www.emergentmind.com/topics/traveling-tournament-problem-ttp