Nicholas (nikaj) (@nikaitsu) header
Nicholas (nikaj) (@nikaitsu) profile picture

Optimizing, AtCoder Heuristic Rank #1, Past: MemSQL(SingleStoreDB)

Joined June 2012
It pains me to say this, I’ve been highest rated in AHC for five years, and I think this change would make it harder for me to stay there, but I believe the only reasonable format for online AHCs going forward is to allow unrestricted AI use. #AHC #AtCoder
1
7
43
4,538
Even with restrictions, AI still shapes which problems organizers set. Cheating will always be hard to detect online, and contestants can’t always know what AI does in background.
1
1
10
1,652
I’d keep onsite rounds human-only and hold as many in-person contests as we can. That’s how we can keep the spirit of the competition alive.
3
11
1,264
#AHC074 This kind of solution might be optimal when all targets are completely random.
1
27
1,380
#mech_39Contest001 Initial greedy chooses high d cells without creating neighbours. SA transitions were mostly random swaps but then I noticed that diagonal adjacency is beneficial and added it.
1
5
707
Had early cutoff over samples with THRESHOLD=5 parameter. If I set this parameter to 0, I get score 187,196,990 which would be 2nd place.
1
2
354
Increasing sample count gets 191,072,902.
2
81
#mech_39Contest001 Nice problem. Simple SA but didn't have time for parameter tuning.
1
13
786
#AHC072 17th provisional. SA on a DAG of operations. Seed 1:
13
726
nikaj took 2nd place in AtCoder Heuristic Contest 071! Performance: 3155 Rating: 3302→3310 (+8) :) #AtCoder #AHC071 atcoder.jp/users/nikaj/histo…
15
623
#AHC071 | 1,590,489 pts | Top 3 Basic SA: The state represents which bricks of size >1 are added. Size=1 bricks are determined in a single top-to-bottom pass in O(H) using bitmasks. Transitions include adding, removing, replacing, or shifting a single brick.
6
1
35
1,608
#AHC070 I could only use two vectors
9
537
First time solving an optimization contest without reading the statement
2
272
#AHC069 Provisional 9th place Use acceptance criteria based on the estimated theta, grid occupancy, and score, checked both before and after finding a placement. In addition to maximizing score, improve the expected availability of adjacent free-cell pairs.
1
10
448
Start candidate regions at random positions and expand them greedily, one cell at a time. When space permits, this produces roughly square regions. Apply a small hill-climbing afterward. Consider placements that relocate zero, one, or two groups.
1
2
264