Ivan, Jun Wei, Ze Chu, Larry
- Variable: Every cell in 9x9 matrix.
- Domain: Every cell, 1-9
- Constraint
- For each col and row, Alldif must be satisfied
- (KIV) Unary constraint for starting puzzle
- Variable ordering
- Most Constrained Variable
- Most Constraining Variable
- Value Ordering
- Least Constraining Value
- Inference Mechanism
- AC3
- Forward Checking
TODO:
- Experiment
- Plot time taken against number of empty cells (for randomly generated puzzles)
- Histogram (frequency against time)
- Variants
- at least 2 variants that has justifications
- Shift AC3 forward
- Clarify most constraining value
- Justification for variant
- Beat benchmark
- AC3 (formally prove that AC3 does well)
- AC3 MRV (see effects of MRV)
- AC3 LCV (see effects of LCV)
- AC3 MRV LCV (see effects of both)
Idea:
- Analytically prove that AC3 works (can also show experimentally when you compare AC3 and FC)
- Experimentaly test w/ MRV and LCV due to its varying effectiveness (ie MRV is more effective in general/stress tests but not in public test cases; LCV not effective alone)
General/Stress tests refer to those in this repository which has implemented the heuristics correctly. Link here. We can confirm the varying effectiveness of MRV and LCV once our experiment set up is up and running.
Explain intuitively backed by evidence why one works better in some cases but not so much in others. (eg puzzles which require diff techniques will work well w/ diff heuristics? But in general, the combination of both usually works) 3. No need to come up with new heuristics - the simple combination will almost beat the benchmark already (missed out input1 by 0.3s).
TODO
- Bug for 1 test case
- Feature extractor