Motivation
aicpp is not only intended to learn mappings from tasks to programs.
A central hypothesis is that training may also modify how efficiently the model explores the program space.
Sequential learning
Tasks are presented sequentially.
After solving one task, the model retains the learned parameters and continues with the next task.
This makes it possible to study whether previous experience reduces the cost of subsequent program synthesis.
Search cost
We measure …
- synthesis time
- number of explored programs
- number of iterations
- cost of the best candidate
- number of programs discovered
- memory usage
Preliminary observations
…