Skip to content

Analysis of 60k Ezreal dataset. This will be used for training the human / superhuman level AI for League of Legends.

License

Notifications You must be signed in to change notification settings

MiscellaneousStuff/tlol-analysis

Repository files navigation

TLoL Dataset Analysis (Patch 13.23: Ezreal 60k)

Assumptions

  • Up to 3 min
    • Minion names are restricted to only those which appear in the first 3 mins (1-hot encoding of their names).
  • Player controlled champ is Ezreal (All of the action spec is strongly tailored around this. Melee champs are really difficult to implement. Ranged champs are easier and if active spell gets correctly implemented will become trivial to implement)

Issues

  • Auto attack inference from missiles is unreliable, especially at 16x replay speed.
    • Not all auto attacks are being correctly scraped (probably needs to be 2x or 4x at most :/). Or activespell needs to be correctly fixed within T_T-Pandoras-Box. This would be far more reliable and allow for much higher throughput.
    • In summary: current detected auto attacks have high true postive accuracy but have low recall.
  • Ezreal Q Spell Usage
    • Discretised aim of the spell is problematic. Only covers +/- 400 units but ezreal q range is 1100 units. This means that the more fine tuned aiming of the spell is just off. This extends also to the W as well.

Conversion Considerations

  • When encoding labels across an entire dataset during bulk conversion for ML datasets, the indexing of the embeddings needs to be consistent. There are two ways to achieve this:
    1. Get the original listing of whatever it is, so for spell names, get every single internal spell name from data dragon for every spell in League of Legends as a list and index into this (This is by far the best method. Need to account for inter-patch changes such as spells being changed or new champions like Hwei which aren't currently accounted for)
    2. Use hasing to compensate by hashing any found spell name, create a build time list of strings for that particular string type (like missile name) and then index into that custom list of strings. This also works but need to be careful during builk conversion run time. As long as it's kept consistent this is also fine and works.

Data Verification Results

  • Auto attack distances between Ezreal and targets are often much higher than AA range, this indicates either a post-processing data handling bug, or something during runtime when scraping? Much more likely to be in the post-process data handling part though.

About

Analysis of 60k Ezreal dataset. This will be used for training the human / superhuman level AI for League of Legends.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published