Figure 7
The figure is divided into three panels labeled (a), (b), and (c) arranged side-by-side. The left panel (a), titled “Hierarchical Learning Curves”, is a line chart with “Accuracy (percent)” on the vertical axis ranging from about 30 to 90 with an interval of 10 percent. The horizontal axis is labeled “Training Matches” and ranges from 0 to 2000 with an interval of 500. Three colored curves represent different decision levels in a hierarchical model: a red line for “Strategic Level”, a blue line for “Tactical Level”, and a green line for” Technical Level”, identified in a legend at the bottom right. All three curves rise steeply from (0, 30) and then flatten, with the strategic curve reaching the highest final accuracy, followed by tactical, then technical. These curves follow the concave-down trend and end between the range of 78.5 and 85 percent at 2000 matches. A vertical dashed grey line at around 300 matches is labeled “Strategic converges”, and a grey line slightly to the right is labeled “Tactical converges”. A black marker on the strategic curve at approximately 500 matches is annotated “80 percent at 500 matches”. The middle panel (b), titled “Graph Structure Evolution”, shows the line graph. The horizontal axis is labeled “Training Epochs” and ranges from 0 to 100 with an interval of 20. The left vertical axis in blue shows “Average Edge Count” and ranges from 40 to 240 with an interval of 20, and the right vertical axis in yellow is labeled “Clustering Coefficient” and ranges from 0.1 to 0.8 with an interval of 0.1. A blue line with circular markers represents “Edge Count”, starting near 50 at epoch 0 and increasing rapidly before gradually approaching around 195 by epoch 100. A yellow line with square markers represents “Clustering”, starting near 0.15 at epoch 0 and rising toward about 0.69 over the same period. A blue text label near the early part of the edge count curve reads “Defensive units emerge”, and a yellow label near the upper part of the clustering curve reads “Attacking triangles form”. The right panel (c), titled “Ablation Study Results”, is a combined bar and line chart. The horizontal axis lists five model variants tilted diagonally: “Full Model”, “w slash o Hierarchy”, “w slash o Cross Att”, “w slash o Adaptive”, and “w slash o I R L”. The left vertical axis is labeled “Accuracy (percent)” and ranges from 60 percent to 95 percent with an interval of 5 percent, and the right vertical axis is labeled “Latency (milliseconds)” and ranges from 20 to 100 with an interval of 10. For each variant, a tall purple bar indicates accuracy, and a black line with circular markers overlaid on the bars indicates latency. The Full Model bar is the highest, labeled “87.8 percent”, with a low latency marker near 28.5 milliseconds near its base. The next model hierarchy reduces accuracy to about 73.1 percent and slightly increases latency. The next model, cross attention, drops accuracy further to about 68.5 percent, with latency increasing to a mid range value. The next model adaptive component yields accuracy around 86.6 percent but a noticeable decrease in latency. The “w slash o I R L” model shows the lowest accuracy, near 65.3 percent, and the high latency of around 86.7 milliseconds. Note: All numerical data values are approximated.Learning curves and sample efficiency