- Estimating Plackett-Luce Parameters using simple neural networks
- Project Year:
2025
- REU Student (s):
Nikol Pushkash | Charles University (Prague, Czech Republic)
- Student 1 Institution:
Charles University (Prague, Czech Republic)
- Project Mentor:
Lirong Xia
- Project Mentor Area:
DIMACS
- Project Abstract:
The Plackett-Luce ranking model is one of the most widely used models for dealing with ranking and preference problems. Although it simplifies work with probability distributions of individual preferences, the parameters are quite complicated to learn from the available preference data. While several methods for doing so are developed and guaranteed to converge over time, some real-world applications require a rather fast but not as precise solution. To fulfill this need, I tried to investigate the possibility of learning Plackett-Luce parameters using simple neural networks and loss functions. This work shows that as the number of voters increases, neural networks are much faster than standard algorithms in determining suboptimal solutions, which are enough for testing and highly loaded services like search engines or non-playable character behavior computation units. These models also introduce a way to initialize less random parameters, which potentially can improve performance of iterative approaches.
