Fix #1252: symmetric split point in dichotomic ObjectiveStrategy

[cprudhom]

Fixes #1252.

Problem

In ObjectiveStrategy.getDecision(), the DICHOTOMIC target was rounded down whatever the resolution policy. Rounding down is right for MINIMIZE (obj <= target), but it is not the mirror image for MAXIMIZE (obj >= target):

• on an interval with an even number of values, the maximization branch kept more than half of them;
• on a two-value interval [a, a+1], the decision obj >= a did not reduce the domain at all.

As a result, minimizing f and maximizing -f did not explore the same tree.

Changes

• Symmetric rounding: the split point is now rounded towards the bound to improve (down when minimizing, up when maximizing), so mirrored objectives try mirrored cuts.
• largerHalfFirst parameter: selects the opposite rounding (up when minimizing, down when maximizing), which tries the larger half of the domain first: a more cautious cut with a stronger refutation. This is option B from the issue discussion. Also available through Search.objectiveStrategy(objective, largerHalfFirst).
• No-op guard: a decision that would not reduce the domain of the objective is no longer returned (getDecision() returns null).
• Single constructor: ObjectiveStrategy(objective, policy, fraction, largerHalfFirst).
• fraction replaces int[] coefs: with DICHOTOMIC, the split point is LB + fraction * (UB - LB), with fraction in [0, 1]. It is ignored for BOTTOM_UP and TOP_DOWN.

Behaviour and API changes

• Search.objectiveStrategy(obj, DICHOTOMIC) tries a different cut when maximizing (rounded up instead of down). Minimization is unchanged by default.
• The two- and three-argument constructors of ObjectiveStrategy are removed. Former coefficients [a, b] become fraction = b / (a + b), e.g. [1, 1] → 0.5, [1, 2] → 2. / 3.
• An out-of-range fraction throws IllegalArgumentException.

Notes

• Math.ceilDiv, suggested in the issue, is only available from Java 18; the rounding is done with Math.floor / Math.ceil on the double split point.
• With a fraction that has no exact binary representation (0.1, 1. / 3), a split point that is theoretically an integer may be shifted by one. 0.5 is exact over the whole int range. This only affects which cut is tried, not the correctness of the search.
• The default (largerHalfFirst = false) is the option proposed in the issue. On the instance reported there, the other option explored fewer nodes; which default is best deserves a wider benchmark.

Tests

• ObjectiveTest#testDichotomicIsSymmetric: minimizing f, maximizing 35 - f and maximizing -f try mirrored cuts, for both values of largerHalfFirst and with negative bounds.
• ObjectiveTest#testDichotomicFraction: mirrored cuts with fractions 0.75 / 0.25, and rejection of an out-of-range fraction.
• DCMST now builds the strategy through Search.objectiveStrategy.

ObjectiveTest and DCMST pass (21 tests), as does mvn checkstyle:check on the solver module. The full test suite was not run locally.

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request