docs/block_optimization_fix.md

Block Optimization Fix Summary

The Problem

NSGA-II optimizer was only producing 33-42 blocks instead of the expected 106 blocks.

Root Causes

1. Reference vs Copy Issue

When storing best solutions from the Pareto front, we stored references instead of copies:

# WRONG - stores references that get overwritten
best_solutions = [(population[i], objectives[i]) for i in fronts[0]]
best_block_solutions = [block_population[i] for i in fronts[0]]

Since population and block_population are replaced each generation with offspring, the stored references pointed to stale/corrupted data.

2. Block-Trainset Mismatch

Even with copies, the stored block assignments were created for a different trainset selection. When the best solution evolved to have different service trainsets, the old block assignment still mapped to old trainset indices.

Example:

• Generation 50: Best solution has trainsets [0, 2, 5] → blocks assigned to indices 0, 2, 5
• Generation 150: Best solution evolves to trainsets [1, 3, 7] → but block assignment still references 0, 2, 5
• Result: Many blocks map to non-service trainsets → lost blocks

The Fix

Always create fresh block assignments for the final best solution:

# Select best solution from Pareto front
if best_solutions:
    best_idx = min(range(len(best_solutions)), 
                  key=lambda i: self.evaluator.fitness_function(best_solutions[i][0]))
    best_solution, best_objectives = best_solutions[best_idx]
    if self.optimize_blocks:
        # Always create fresh block assignment for the best solution
        # to ensure all 106 blocks are properly assigned
        best_block_sol = self._create_block_assignment(best_solution)

The _create_block_assignment distributes all blocks evenly across current service trainsets:

def _create_block_assignment(self, trainset_sol: np.ndarray) -> np.ndarray:
    service_indices = np.where(trainset_sol == 0)[0]
    
    if len(service_indices) == 0:
        return np.full(self.n_blocks, -1, dtype=int)
    
    # Distribute blocks evenly across service trains
    block_sol = np.zeros(self.n_blocks, dtype=int)
    for i in range(self.n_blocks):
        block_sol[i] = service_indices[i % len(service_indices)]
    
    return block_sol

Result

Optimizer	Before Fix	After Fix
GA	106 ✓	106 ✓
CMA-ES	106 ✓	106 ✓
PSO	106 ✓	106 ✓
SA	106 ✓	106 ✓
NSGA-II	33-42 ✗	106 ✓

All optimizers now correctly assign all 106 service blocks.