Experiment 2: Towers of Hanoi
3 pegs: S, A, D; n disks on S; move all to D, but never a larger on a smaller. Additional primitives: movdiskSD(); exchSA(); exchAD().
Fastest solution costs 2n-1 moves.
Anderson 1986: reinforcement learning, nɜ.
Langley 1985: production systems, nɞ.
Baum & Durdanovic 1999: simpler blocks problem scales linearly, nɞ (Kwee 2001)
Nonlearning AI planners: n size < 100,000 (because they just search in raw solution space!)
OOPS: n ? 30; solution size > 1,000,000,000 (because OOPS searches in space of solution-computing programs!)
Speedup: first learn seemingly unrelated language task 1n2n!
This gives us an opportunity to demonstrate incremental learning - knowledge transfer from one task to the next
Back to J. Schmidhuber's OOPS page