Sandbox idea
Step 1. Write a SW that records and remembers movement.
Step 2. Have some competent players run a tile (f.ex TL, jungle gym etc) 1v1 until the survivor is downed, and repeat for a sufficient number of times.
Step 3. Repeat for all killers and all generic tiles.
Step 4. Consolidate interaction data from "Step 1." in the form: when survivor moved from such to such coordinates and/or performed such and such vault or pallet drop, the killer responded by moving between such and such points and/or swinging and/or lunging and/or using ability. If pallet still undropped, unbroken. If survivor injured, if not yet injured, if bloodlust level so and so; and so on and so forth. And vice versa for if killer did, then survivor did.
Step 5. Assign weight to these remembered interactions according to what led to successfully downing quickly and conversely staying up for a long time, and under which conditions. If different combinations of survivor and killer moves and actions led to a down or vice versa under the same conditions, then give the most frequently successful ones a higher weight.
Step 6. Use this data as your basis for a looping AI for both survivor and killer.
Step 7. Create a tiny map and implement it into the game, in which we can run one generic tile at a time against the AI for practice, as either killer or survivor. Perhaps the killer can m1 a "gong" in the map to render a different tile, and the survivor can flip a switch for the same purpose, and another switch for a different killer, a third switch to reset pallets etc etc.
Bonus suggestion: Invite Fog whisperers to do the running. This way, all of their fans get the chance to go against that player in a chase (sort of).
TL/DR: Since it's so hard to write a competent AI decision-tree for an entire match, don't! Instead, record the movements of actual players tile by tile, and let that record form the decision-pattern for an AI looping minigame. AI modes are mostly good for grinding mechanical skill and learning the basics anyway, and not so much for strategy (even in games that DO offer a full-game AI mode).
Not only would this allow an AI to use advanced tactics like 360-ing or moonwalking without actually having to understand what it's doing (it would simply use its bank of prerecorded movement patterns) but it would also be able to be unpredictable (by randomizing between recorded movements) and competent (by giving the randomness a bias towards more successful records).
Comments
-
I think there will be still some issues to solve to make that but sounds like a good way to implement an actual training mode and a decent tutorial. For right now they need to focus on other things for sure but I hope they consider this if they have the people to do that work
0 -
Yh, cant wait for an AI to teabag me.
4 -
This would be fantastic to add, but it would also be a way easier said than done. Recording movements for every single tile and every possible way of running the tile is one thing on its own, but then you need to consider tile variations (where did the pallet spawn, where is the window), how to run each variation against different killers (you wouldn't loop a clown, much like how you wouldn't download a car).
Beyond all that though, there's a couple big issues. Either A) The AI can read your every move through LOS blockers and is able to react unrealistically, or B) The AI is a big dumb and sticks to their path no matter what, making it easy to abuse. Then you've got to take each killer and perk into consideration. What's stopping me from playing trapper and blatantly putting a trap somewhere along their preset loop? What if I play Spirit, how is the AI going to know how to mindgame me? What if I'm using coulrophobia, are they going to be smart enough to stop healing or are they going to keep wasting their time? Or if I've got BBQ & Chili, are they going to hide their auras?
It would be totally awesome to have an AI in Dead by Daylight, and it would be particularly great if that AI could take over if someone loses connection. But since the game is procedurally generated and chases are largely based on mindgames and tiny optimizations, making an AI that can compete with a real player is a massive task.
9 -
I'm aware that I was more spitballing ideas than presenting solutions.
I don't know why I mentined all killers, as many of them do indeed have powers that make you not want to commit to looping; that's true.
The most useful ones would probably be:
Pig (learning when and how to use and how to vs her power in a loop).
Billy (when and how to curve with m2, and how to avoid it. When to settle for the m1).
Huntress (when you can crouch to avoid hatchets and when she can find an angle, the right timing to fake a vault etc).
And maybe Wraith for the generic m1 match-up.
I was thinking perkless and addon-less, and I suppose every tile variation doesn't have to be recorded either. As long as there are a few different scenarios available to practice, I think it would be a relief for those who don't have good KYF opportunities, though. A new Billy trying to learn how to curve around a car in public matches can be a dragged-out and humiliating experience (and possibly a personal experience of yours truly).
And let's be honest. It takes countless repetitions to learn how to consistently fast-vault in a confined space, I'm sure we've all encountered people who spend an entire public match just practising their moves (or indeed seen it on stream, no names).
Hmmm...Maybe what we really need is an obstacle course. Like the police academy, but with more blood. Or even just letting us solo-explore maps and run around might go a long way. I guess I'm still just spitballing...
PS. How do you know I don't possess a prototype futuristic 3D-printer assembly line which allows me to actually download a car? The gall of these preconceptions!
0
