Further racing AI refinements

[Alayan-stk-2]

This is a list of the main areas where the AI could use significant improvements with thoughts on each. Consider this as a sort of progress-tracking list.

To identify issues, I highly recommend launching a race and set the camera to one of the AI kart. This allows to much better see how the AI acts.

• [ ] 1. Yet more item-handling improvements. #3143 should improve things a lot here. There are still additional possibilities (a lot of them mentioned in code comments at appropriate places). Low priority as what's still possible to do won't improve race performance as much.
• [ ] 2. Better pathing. The AI relies heavily on drivequads. This means that : a)it will never consider to go outside of the driveline even when it would be beneficial (lighthouse, I'm looking at you) ; b)when the road do a kind of ~, it will largely follow the full curves instead of trying a more straight path. While (a) may be judged to be a track design issue ; (b) is one of the biggest current AI weaknesses
• [ ] 3. Better skidding. The main issues being : the AI often don't try enough to use it in slight curves ; the AI doesn't know it's better to keep skidding a bit to reap the bonus ; the AI tries to skid for too long in big curves (it's often more efficient to skid, let the boost activate, skid again).
• [x] 4. Action under plunger influence. Currently, it makes the kart slow down to half-speed. Modulating it according to difficulty (and maybe also influence item collection/avoidance !) may allow it to be a bit less affected in higher difficulties
• [ ] 5. Going outside of track in narrow paths. Self-explanatory. See the old mine, fort magma (with the bridge), green valley, probably others.
• [ ] 6. Turning too sharply to try to catch an item/touch a kart. Just go watch the AI trying to take the nitro on the right side a bit before the road passes under the bridge in the Old Mine and you will understand. What's needed : a better angle sorting so that the AI don't consider it doable to touch the item. For karts, it's more complicated as relative speed is also involved.
• [ ] 7. Conflicting/inconsistent code-paths. The AI uses quite a bunch of code, and as some part try to handle some specific case, they may contradict or ignore what other paths do. For example, nitro being used when AI wants to give a bomb back shortcuts the whole usual nitro handling.
• [ ] 8. Teach the AI how to get back in a correct position when it ends oriented sideways or backwards. In those situations, it advances very slowly and seems clueless as to how reorient itself. It would be much more quick to either rescue the kart (not very clean though) or to go backwards while turning before advancing and turning in the other direction (a trajectory tracing a kind of curved Y)
• [ ] 9. Improve bad items avoidance. Currently, when trying to get a good item, the AI check if there is not a bad item in its way. But it doesn't check if there is not a bad item just after the good item (see the banana/nitro/banana in Cornfield Crossing ; it also often appears randomly with gums)

There may be others. Write a comment if you think of one.

[Alayan-stk-2]

Seen looking at AI races : the AI sometimes insists on keeping skidding so much, while it already could get the bonus and there is a straight line ahead, that it goes off-track. E.g. ; at the last curve in XR591 ; I got an AI falling to the left and having to go do the curve again... twice in a row !!!

[Alayan-stk-2]

For (6) ; here is a helpful image to better understand the issue :

Black is how AI seems to evaluate things ; blue is how it may drive. (Not exact values ; it's only the general idea)

This means that : (a) it will miss some items it thought it could catch (those who fall between the blue and black line) ; (b) if the item is a bit further and can be catched (beyond the black/blue crossing but close to the black line) it will arrive at it with a much worse direction than expected, throwing it out of track.

[Alayan-stk-2]

The new pathing algorithm was supposed to look ahead at left and right points until the line they trace join. However, it is bugged.

        // Test if the next left point is to the right of the left
        // line. If so, a new left line is defined.

If the next left point was instead to the left of the current left line (or the right to the right of the current right line), then we would break out of the loop, which defeats the whole point of trying to get the left and right line to join.

Then, the other issue is that it's the center of the last node in the loop which is used, rather than the appropriate left/right point matching the joining lines.

With that fixed, the pathing algorithm should work much better.

Then, it is a matter of making skidding work nicely.

EDIT : With those fixes applied, the pathing is finally good ! A tux completed 3 time-trial lighthouse laps in 1:30.04 without skidding at all. For comparison purposes, I got 1:23.48 in those conditions (7% less) but that is getting some nitros the AI didn't and better managing it/zippers. If I miss the double nitro at one place in the path like the AI (which points to a weakness in the item collection algorithm), I lose 3-4 seconds. The previous AI steer more to the center and get those cans, but (without skidding) only beat the new time by 1,5 seconds ; so there is around 2 seconds in pathing difference.