Manish.
Staff Member
- Joined
- Aug 22, 2016
- Location
- Vijayawada, India.
- Profile Flag
- India
- Online Cricket Games Owned
-
- Don Bradman Cricket 14 - Steam PC
I am excited to share details on three AI models I have developed to simulate cricket matches. These models aim to realistically simulate decisions made during a match, like which bowler to select and which batsman to send in next.
The first model focuses on bowler selection. It analyzes the current match situation - opposition score, wickets in hand, and batsmen at the crease - and chooses the optimal bowler from the playing XI using a random forest classifier. This model has achieved 96% accuracy on test data.
The second model handles batsman selection when a wicket falls. It looks at the team's current score, wickets in hand, and remaining batsmen to decide whether a power hitter or strike rotator should come in next. A 92% accuracy has been reached using random forest classification.
The third model ties the first two together to actually simulate full matches ball-by-ball. Using extensive T20 data and models like XGBoost and neural networks, it tries to replicate the endless possibilities of a cricket game. While accuracy is lower at 56% for XGBoost and 49% for neural networks, I believe this is strong given the problem complexity and data used.
Current limitations are the focus on T20 cricket and reliance on real players from historical matches. In the future, I hope to expand to ODI simulation and enable custom players. But for now, I'm working on finalizing the first complete match simulations. Expect to see some sample output within the next week or two!
The first model focuses on bowler selection. It analyzes the current match situation - opposition score, wickets in hand, and batsmen at the crease - and chooses the optimal bowler from the playing XI using a random forest classifier. This model has achieved 96% accuracy on test data.
The second model handles batsman selection when a wicket falls. It looks at the team's current score, wickets in hand, and remaining batsmen to decide whether a power hitter or strike rotator should come in next. A 92% accuracy has been reached using random forest classification.
The third model ties the first two together to actually simulate full matches ball-by-ball. Using extensive T20 data and models like XGBoost and neural networks, it tries to replicate the endless possibilities of a cricket game. While accuracy is lower at 56% for XGBoost and 49% for neural networks, I believe this is strong given the problem complexity and data used.
Current limitations are the focus on T20 cricket and reliance on real players from historical matches. In the future, I hope to expand to ODI simulation and enable custom players. But for now, I'm working on finalizing the first complete match simulations. Expect to see some sample output within the next week or two!
