batteryswapaichallenge/YassYBS26Sol43

Sol43: Zero-Cost Merge Planner

This submission implements a major structural upgrade to the planner's dynamic programming logic, building directly on the success of Sol41.

Architecture

The core insight of this solution is that the public dataset severely punishes Overtime and Weekly Limit breaches. When multiple batteries fail in the same time window, a naive planner will schedule multiple separate trips, rapidly stacking overtime and weekly penalties.

Sol41 attempted to solve this by aggressively spreading trips out across weeks (via week_pen_inflate = 1.5 and min_day_span = 14), but it did this blindly. It often just moved the congestion to another week, or created brand new 20-hour trips on days that were otherwise empty.

Sol43 introduces Zero-Cost Trip Merging. The solver's internal week_relief local search mechanism has been rewritten to actively scan the 14-day surrounding window for days that already have a trip scheduled to the exact same building.

When it finds a matching building, it merges the batteries into that existing trip. This effectively deletes 20 hours of logistics travel time from the schedule entirely, dodging the weekly limit while simultaneously crashing the daily overtime penalties.

Additionally, this version enables weekly_aware_greedy=True, ensuring the initial schedule construction is mindful of the 25-hour weekly limit before the local search even begins.

Local Evaluation

Local evaluation metrics (especially early and late swap penalties) are not perfectly correlated with leaderboard performance due to the 16-swap evaluation cap and the extreme density of the leaderboard dataset. However, using the exact local validation method from Sol41, this approach successfully dropped the local Weekly Limit penalty from 41.67 down to 25.0 (which equates to roughly 0.25 weekly breaches per scenario).

Usage

This package contains the pre-trained thermal_ranker ensemble and the custom RobustHazardPlanner. No further training is required. Simply submit the package to the evaluation environment.