Evolutionary optimisation,
with receipts.

01

Double-blind review · title withheld until decision

With

Jonathan Mwaura

GA-tuned ensemble agent for Battleship

Four Battleship strategies, one heatmap each, blended by weights a genetic algorithm learned. Averages 40.5 moves to win and beats the 2011 benchmark.

40.5

avg moves to win

42

previous benchmark (Berry, 2011)

−4

moves vs uniform-weight ensemble

4

strategies blended

Abstract

Existing Battleship agents commit to one technique: probability density functions, Monte Carlo sampling, neural networks or heuristics. None dominates. This paper proposes an ensemble in which four strategies (a per-ship probability density function, a convolutional neural network, Monte Carlo occupancy sampling and an information-gain heuristic) each produce a heatmap over the board, fused by a sequential, scale-preserving blending rule. A real-coded genetic algorithm learns the blend weights against a fitness of win rate and moves-to-win. The resulting agent averages 40.5 moves, outperforms every single-strategy component, improves on a uniform-weight ensemble with the identical architecture by about 4 moves, and compares favourably with the strongest published benchmarks of 42 to 45 moves. The learned weights are interpretable and the GA pruned one redundant strategy to zero.

02

Double-blind review · title withheld until decision

With

Jonathan Mwaura

Adaptive generational ↔ steady-state switching for GAs

A genetic algorithm that starts generational, watches its own elite centroid, and switches itself to steady-state the moment exploration stalls. No hand-picked switch point.

60 / 60

runs where the trigger fired

20 / 20

successes on every benchmark

0.9986

mean accuracy on Wine

2% · 10%

centroid & variance thresholds

Abstract

Generational GAs replace the whole population each iteration and explore broadly; steady-state GAs replace one or two individuals and exploit aggressively. Neither wins everywhere, and fixed-schedule hybrids make you choose the switch point by hand. DEGA is a hybrid framework that combines both regimes, and Method 2 is a centroid-stagnation trigger for a one-way generational-to-steady-state transition. It fires when the 10-iteration mean of elite-centroid displacement drops below 2% of its first non-zero value and elite fitness variance drops below 10% of its initial value, after a minimum generational duration. Over 20 matched runs on three neural-network weight-optimisation problems (XOR, Iris, Wine) against pure GGA, pure SSGA and six fixed-schedule variants, DynDEGA-M2 succeeds 20/20 on every benchmark, with mean validation accuracy of 0.9967 on Iris and 0.9986 on Wine, and lands within one standard deviation of the best fixed schedule every time. The trigger fired in all 60 runs, in the first 7.4% to 26.6% of the budget.