Quoted from HighVoltage:
There's various ways to approach this with AI.
HighVoltage is essentially correct. Possible but very, very, very difficult. The model would take a lot of time to develop and you would need to a lot of data from the environment; score, states of lights & switches, and ball position, velocity and angle. Plus to have the same chance as a good player, the computer would have to be able to physically nudge the machine.
I would think you would need an ANN algorithm over over supervised machine learning. Machine learning would work for something like the Phantom Flip on Monster Bash - but not a complete game that can navigate a ruleset and maximize scoring.
Hmm... fun to think about.