Microsoft’s attempt to use machine learning to improve on the Duckworth-Lewis method in cricket has been dismissed by the current custodian of the system.
For the uninitiated, the Duckworth-Lewis method, or D/L method for short, is used to calculate the score that the team batting second in a limited overs cricket match needs to reach if the match is affected by rain.
For example, if the team batting first scores 400 in 50 overs for five wickets, but rain reduces the second team’s innings to 40 overs, the D/L method may put forward a score of 300.
It was invented by statisticians Frank Duckworth and Tony Lewis and was first used in 1997 in a match between England and Zimbabwe. It is now regularly used at matches at all levels.
The system is sold to clubs in Standard and Professional editions.
It is now overseen by professor Steven Stern of the Queensland University of Technology, and has been officially renamed as the Duckworth-Lewis-Stern method, or DLS method.
The system has been in use for many years, but it is not without criticism, such as favouring wickets too heavily and being too complex for the average fan to understand.
So, in an attempt to use technology to improve the system, Sarvashrestha Paliwal, Azure business lead for Microsoft India, said that machine learning could improve the system.
“The D/L table is static and does not take into consideration the latest game statistics (e.g. which teams are playing better this season, ranking of players, etc),” he said.
“We believe we can use historical Twenty20 data to derive an always up-to-date D/L table that takes into account these latest statistics. This can be operationalised using Azure machine learning and run on a frequent basis to always produce an updated D/L table.”
Paliwal explained that the team used data from Twenty20s to capture around 153,000 rows of information from matches to create a new model for predicting second innings scores.
“Using a Jupyter notebook, we show the data exploration on how we can derive a better D/L table by applying quadratic curve-fitting with constraints techniques using the T/20 data,” he said.
“This Jupyter notebook is now available to the data science and cricket communities so they can take this foundational information and work together to improve the state of cricket analytics.”
However, professor Stern had a few things to say about this, pointing out some errors in the assumptions made and the rationale for some elements in the system not considered by Microsoft.
Firstly, he explained that, contrary to the claims by Paliwal, the DLS is not static but “regularly updated”.
"Generally, I re-analyse data annually and a new version of the method is produced biannually,” he said.
Secondly, Stern had a problem with the idea of using performance trends in the system.
“Doesn't the better team in a contest already have sufficient advantage? Codifying their advantage into the rain rule seems rather harsh on the weaker opponent,” he said.
Lastly, Stern wrote that taking player abilities into account could also allow teams to “game the system” depending on who was at the crease and who was left to bat.
“For example, if a rain rule target depends on who is at the crease at the time of interruption, it may well be to a team's advantage to have their best batsman at the crease and if he/she is not already there, there suddenly is an incentive to intentionally lose a wicket (and, conversely, for the fielding side to not take a wicket),” he explained.
Stern ended by saying that he will happily talk to the team at Microsoft to discuss this further, so there could still be a part for machine learning to play in cricket. But for now, the good old Duckworth-Lewis-Stern system remains top of the order.
RISC OS 5 to form the basis of RISC OS Open after Castle Technology sells to RISC OS Developments
A smartphone maker fiddling its benchmarking scores? That's unusual, isn't it?
'We are making good progress on 10nm,' claims Intel
Engineer calculates that Chengdu's plan to replace streetlights with artificial moonlight would cost $100bn