I'm slightly familiar with the reinforcement learning (through this edx.org MOOC). Also I'm aware of reactive planning technique. I want to write a vehicle bot for a racing game, where cars have to complete laps and avoid collisions.
Is it sensible to use reinforcement learning for steering? Or should I stick with reactive planning?
In Artificial Intelligence for Games 2E the authors discuss racing games regularly and say that approaches built on steering behavior primitives (i.e. reactive planning) can be very well-suited to the domain. (I'm in the middle of reading it now, so I can't say whether they make the same claims for reinforcement learning.)
It of course depends on your exact needs, but I believe reinforcement learning is likely to be considerably more engineering effort than you need. You'll also have fewer resources available online for this particular application of reinforcement learning.
Of course, if you simply want to cut your teeth on reinforcement learning, it's certainly a possible pathway.
If you decide to stick with steering behaviors, you may find the source code accompanying the above-mentioned book a useful reference (and of course, the book itself). The steering-behavior-specific demos are the c03_* folders under https://github.com/idmillington/aicore/tree/master/src/demos. I haven't looked over the source code yet, so I can't say how accessible it would be without the book at hand.
