EVE Fanfest 2018: CCP Games has big plans for data collection and machine learning
The feature will be delivered as a new Activity Tracker window in the game client that will show players detailed stats on almost everything they’ve done in-game since the tracker went live. This in itself is useful, both for helping players set goals and for highlighting other areas of the game they might not have given a fair shake yet and so might enjoy. Behind the scenes, the way that CCP is collecting this detailed data and the implications of its use are really fascinating, and there are even plans to use machine learning to look for patterns in this data that would help identify bots.
Read on for a breakdown of exactly how masses of new data is being captured on EVE players, and how it could be put to use in the future.
Actions in EVE Online are essentially transactions in a database, and even a supercomputer cluster such as the one the EVE server runs on has performance limitations when operating on a database. These constraints have forced CCP to be very careful with what information they keep in order to keep the server running smoothly, and anything that isn’t needed for a specific purpose has to discarded. Executive Producer Andie ‘CCP Seagull’ Nordgren explained to me that this model has now changed:
“For legacy performance reasons, the server is pretty harsh at culling what gets saved to the database. … Now we’re pushing data out to message queue instead and just sticking it in the cloud, basically.” – Andie Nordgren
This means CCP can now record an insane granularity of detail about what every inhabitant of EVE is getting up to. When a ship is blown up in-game, they’ll be able to record each bit of damage done to that player, all of the remote assistance done to the attackers, even how many kill-marks each ship had and what skin it was using. Players will enjoy the fruits of this level of detail through the new Activity Tracker window, advancements in third party tools, and highly requested features such as logistics pilots finally being credited on killmails. Many of these features just weren’t feasible before because the data required was always discarded, and the possibilities this opens up are pretty awesome.
The biggest payoff from collecting data en masse from players like this will of course be for CCP, as it will allow developers to more easily confirm their hypotheses about how players engage with the game. CCP already leverages this kind of data discovery as part of the design process, but the studio will now have access to a lot more information on what players actually do in-game.
“We look at a lot of data all over the place, and I think it’s more interesting to me to be hypothesis-driven, and then you use the data. The data never drives anything, you have some hypothesis that you might use data to validate or test your assumptions, but the drivers for design should be creative.” – Andie Nordgren
What’s all this about AI and botting?
Perhaps the most interesting possibility to me personally is what you can do with a massive collection of coherent and verifiable data today — feed it into a machine learning algorithm. Machine learning artificial intelligence systems are incredibly good at classifying things into categories accurately as long as they have huge databases to work with and plenty of examples of pre-classified results. They can often outperform a human at the same task and find strong correlations and predictors in the data that a human wouldn’t even think to look for.
“There’s certainly the potential there where I think we will have all new ways to operate on that data. It’s like surveillance obviously, but of course when you’re operating in a virtual space there’s no such thing as not being surveilled in one form or another. But this is certainly an increased granularity on understanding what you do.” – Andie Nordgren
Once enough data are collected, CCP could potentially set up a machine learning system with the data from known botters that are caught by the GM team, and it should slowly learn to identify botters by patterns in their detailed activity data. The AI would then be capable of finding bots in the live game even if no players report them, and could then flag them up for investigation by human GMs. This isn’t something that CCP will be doing right away as the new data collection tools will need to be “released and established first.” But with the new data channels going live on May 29th with the Into the Abyss expansion, botters’ days in EVE Online may very well be numbered.