The stats of Valorant pro players might be able to be used to guess the agent of a generic player. I made a web app using React and Spring Boot with a machine learning algorithm powered by scikit-learn and the gaussian naive bayes categorization algorithm.
Initially I tried using the absolute best players (n=~250) as a sample size, but this was not enough to train my model. So I re-ran my scraper to allow for any pro play in VCT2022. However, this results in a decrease of quality of results. So I decided to instead take the results from multiple circuits (linked in Resources) with the following parameters:
- At least 500 opponent rating
- At least 100 rounds played
Since franchising resulted in less T1 games being played, I was unable to successfully fully update the set. I added the 2023 VCT data (which did not significantly change accuracy levels) in hopes of predicting gekko better. I lowered the threshold of maps for more recent data to add emphasis and increase the gekko representation in the dataset. This resulted in a small accuracy decrease.
- Not all agents receive equal play. This means it is harder to correlate data with lesser played agents
- A normal player might have very different stats from a pro
- Some agents have similar roles, so differentiating between them from stats is inherently challenging. To remedy this, on the frontend I display the agent and the role as guesses.
- A constantly changing meta means this might not be as accurate on modern data
I made three different models. The model used in the API is the agent guessing model. This model has an accuracy of ~43%. Considering there are 20 agents and some of those agents are bound to have similar stats, I would characterize this as extremely successul, especially given the restrictions on the size of the data.
However, I wondered how accurate it was accounting for agent similarity, so I made 2 other models.
The first of these tries to categorize by the agent class (Duelist/Initiator/Controller/Sentinel). This is the riot-sanctioned categories that I could categorize into. With this in mind, my accuracy was ~61%. This is a great improvement, but I knew I could do better.
There are some agents which when played in pro-play, fit better into other classes. Therefore, I made my own 4 roles (Star Player/Assistant to the Star Player/Smokes Specialist/Area Denier). This moved the following agents from their default role:
- Chamber from Area Denier to Star Player
- Viper from Smokes Specialist to Area Denier
- Sage from Area Denier to Assistant to the Star Player
| Star Player | Assistant to the Star Player | Smokes Specialist | Area Denier |
|---|---|---|---|
| Chamber | Breach | Astra | Cypher |
| Jett | Fade | Brimstone | Killjoy |
| Phoenix | Kayo | Harbor | Viper |
| Raze | Sage | Omen | |
| Reyna | Skye | ||
| Yoru | Sova |
When categorizing like this, I achieved a ~70% success rate using SVC.
In the end, I send the agent guess to the frontend and then manually categorize, but I am very happy with the accuracy of my results given the variety and similarity of agents.
There are 3 main folders for this project:
- frontend
- backend
- data
The first two are fairly self-explanatary, but the third is where I did web-scraping and machine-learning focused work.
I scraped my dataset from vlr.gg where I took a sample of the top players in 2022.
Specific links were taken from:
I also want to credit tkomarlu for helping me plan out the project/learn how to apply scikit-learn to the idea.
- .env on frontend needs REACT_APP_BASE_URL set to BE url
- build the files and export them into out. The PMML for your model must sit next to your .jar file (see example in repo)
- Run Dockerfile to create a website image
I decided not to test this project since each component is very small/simple. If I end up adding to it, I may add testing, but for now I am more concerned with it working as I expect it to.