Utility issues with Graphplan Memos
EBL strategies typically suffer from significant utility problems
- Cost of storing no-goods; Cost of matching the no-goods
- Solver needs to selectively forget learned no-goods
- (size-based learning; relevance based learning etc.)
- Why is this not a significant issue with GP+EBL?
- Reason: Memos correspond to a very conservative form of no-good learning