How science can improve sporting performance

Table ten­nis, a pop­u­lar rack­et sport, has been an Olympic dis­ci­pline since 1988. It is also a sub­ject for aca­d­e­m­ic study in which uni­ver­si­ty researchers analyse the pat­terns and tac­tics in the game to improve play­ers’ com­pet­i­tive per­for­mance. Researchers at the École Cen­trale Lyon’s Lab­o­ra­toire d’in­for­ma­tique en images et sys­tèmes d’in­for­ma­tion (LIRIS UMR 5205 CNRS) are work­ing in close col­lab­o­ra­tion with the French Table Ten­nis Fed­er­a­tion (FFTT), the body respon­si­ble for man­ag­ing table ten­nis in France and pro­vid­ing tech­ni­cal sup­port to the high­est-lev­el play­ers. The sci­en­tists are analysing videos of match­es avail­able on broad­cast­ing plat­forms online and attempt­ing to iden­ti­fy pat­terns of play based on datasets con­tain­ing rel­a­tive­ly short (four to five strokes on aver­age) but com­plex (some twen­ty descrip­tors per stroke) sequences of rack­et strokes. They have their work cut out in the run-up to this sum­mer’s 2024 Olympic Games in Paris. 

In a table ten­nis match, play­ers take turns hit­ting the ball with their rack­et and bounc­ing it off the oppo­nen­t’s side of the table – except when serv­ing, when the ball must bounce off both sides. A ral­ly is lost if a play­er fails to return the ball in accor­dance with these rules. A play­er wins a set when 11 points or more are reached, with a dif­fer­ence of two points between the opponents.

In their work, the researchers led by Romain Vuille­mot, analyse the posi­tion of the play­ers, the way they move dur­ing a stroke, and the move­ment and tra­jec­to­ry of the ball. The sci­en­tists are focus­ing on analysing win­ning stroke com­bi­na­tions to char­ac­terise play­er tac­tics. Accord­ing to FFTT coach­es, a tac­tic con­sists of two con­sec­u­tive strokes for one of the play­ers, which means that in a ral­ly, it is played in three con­sec­u­tive strokes for both play­ers. The serv­ing play­er con­trols the first two strokes of the game, then hits the ball back to the oppo­nent in a way to poten­tial­ly win the point. Exchanges beyond the first three strokes are inter­est­ing tech­ni­cal­ly and can take prece­dence over tactics.

Find­ing use­ful tac­tics in this way is not lim­it­ed to table ten­nis com­pe­ti­tions but can also be applied to oth­er sports such as foot­ball and box­ing, even though these involve much more phys­i­cal con­tact between play­ers. In foot­ball, how­ev­er, a sequence is defined as a list of sev­er­al con­sec­u­tive moves made by the same play­er or moves by dif­fer­ent play­ers at the same time. In a table ten­nis match, the sequences are actu­al­ly a ral­ly (and there­fore high­ly syn­chro­nous). Thus, in a series of con­sec­u­tive table ten­nis shots, the two play­ers appear alter­nate­ly in a sequence. This means that there is a strong cor­re­spon­dence between the sequence of strokes and the play­er, with actions/reactions, antic­i­pa­tions, and dominations.

The researchers analysed the sequence of strokes in a ral­ly until it was won by one of the two play­ers. A sequence has the fol­low­ing structure:

• A serve (from the right or left side) that hits one of the nine pos­si­ble impact zones on the oppo­nen­t’s side of the table.

• A sequence of strokes described by the type of stroke (con­trol, attack, or push), whether it is a back­hand or fore­hand stroke and the impact zone of the ball.

Oth­er descrip­tors are also analysed, such as spin, tech­nique, and play­er position.

Using a new algo­rithm that they devel­oped in their lab­o­ra­to­ry, the researchers con­struct­ed a graph rep­re­sent­ing a set of ral­lies based on visu­al data from a video. The nodes of the graph rep­re­sent the moves, and its edges rep­re­sent the tran­si­tions between moves. The nodes are ordered so that ral­lies are “read” from left to right: the left­most node is the serve, and the right­most node is the win­ner of the rally. 

“By analysing the move­ments of each play­er, the move­ments of the rack­et and the tra­jec­to­ries of the ball on the table, we can clas­si­fy the type of stroke into dif­fer­ent cat­e­gories and analyse what hap­pens dur­ing each point,” explains Romain Vuille­mot. “We then try to under­stand the strat­e­gy behind a giv­en stroke. To do this, we need to under­stand what the play­er is doing in gen­er­al and how his oppo­nent is react­ing, that is, has he adapt­ed to his com­peti­tor’s game? This is a very com­plex task, but our graphs already allow us to iden­ti­fy poten­tial­ly effec­tive sequences. We then need to con­tex­tu­alise these results, par­tic­u­lar­ly in terms of the score or the dom­i­nance of one of the play­ers”. Tac­ti­cal­ly inter­est­ing points are often few and far between, but they are often deci­sive in explain­ing a victory. 

The results of the analy­ses will pro­vide play­ers, and their coach­es, with the infor­ma­tion they need to under­stand and adapt to the game of their oppo­nents, and thus improve their own strategies.

“What we have found is that the tac­tics adopt­ed depend on each play­er and their style of play,” explains Aymer­ic Erades, a doc­tor­al stu­dent work­ing on the project with Romain Vuille­mot. “What’s inter­est­ing is that a giv­en play­er won’t always play in the same way. The idea is there­fore to detect all the pos­si­ble play com­bi­na­tions and then to under­stand the ones that will be used dur­ing the game, par­tic­u­lar­ly against an oppo­nent whose strengths, weak­ness­es, and gen­er­al style of play we are famil­iar with.” 

To date, the researchers have analysed around 30 match­es in their entire­ty and are in the process of analysing a fur­ther 70. “This num­ber could increase in the run up to the Olympic Games this sum­mer since table ten­nis play­ers are busy get­ting ready for their match­es and coach­es are ask­ing us to pro­vide them with addi­tion­al data on new oppo­nents,” explains Aymer­ic Erades. 

“Between now and the sum­mer, we hope to have obtained tac­ti­cal analy­ses of all the poten­tial oppo­nents that the French team is like­ly to encounter to under­stand how they play and ulti­mate­ly com­mu­ni­cate this infor­ma­tion to the fed­er­a­tion so that our play­ers have the best chances of beat­ing their oppo­nents,” adds Romain Vuillemot. 

Isabelle Dumé

Ref­er­ences:

Explor­ing Table Ten­nis Ana­lyt­ics: Dom­i­na­tion, Expect­ed Score and Shot Diver­si­ty. Machine Learn­ing and Data Min­ing for Sports Ana­lyt­ics work­shop MLSA, 2023

Visu­al Analy­sis of Table Ten­nis Game Tac­tics. Journée Visu 2023, 22 juin 2023, Saclay (France)

Pierre Dulu­ard, Xin­qing Li, Marc Plante­vit, Céline Robardet, Romain Vuille­mot. Dis­cov­er­ing and Visu­al­iz­ing Tac­tics in a Table Ten­nis Game Based on Sub­group Dis­cov­ery. ECML/PKDD 2022 Work­shop, Greno­ble, France, 2022. hal-03768114