Play-Persona: Modeling Player Behaviour in Computer Games
Play-Persona: Modeling Player Behaviour in Computer Games
Play-Persona: Modeling Player Behaviour in Computer Games
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
weight values we repeat the tra<strong>in</strong><strong>in</strong>g 20 times — us<strong>in</strong>g dissimilar <strong>in</strong>itial weight vectors — and select<br />
the ESOM with the smallest average quantization error (see Section VI). The highest-perform<strong>in</strong>g<br />
ESOM that is exam<strong>in</strong>ed <strong>in</strong> the rema<strong>in</strong>ed of this paper has a quantization error of 0.038 and a<br />
correspond<strong>in</strong>g topographic error of 0.005. As a basel<strong>in</strong>e performance to compare aga<strong>in</strong>st, the mean<br />
quantization and topographic error for 10 randomly generated ESOMs equals 0.1744 and 0.9983<br />
respectively.<br />
2) ESOM visualization: The tra<strong>in</strong><strong>in</strong>g data can be clustered by observation of the best<br />
perform<strong>in</strong>g ESOM. The U-matrix depicted <strong>in</strong> Fig. 5(a) is a visualization of the local distance<br />
structure <strong>in</strong> the data placed onto the two-dimensional map. The average distance value between<br />
each neuron’s weight vector and the weight vectors of its immediate neighbors corresponds to the<br />
height of that neuron <strong>in</strong> the U-matrix (positioned at the map coord<strong>in</strong>ates of the neuron). Thus,<br />
Umatrix values are large <strong>in</strong> areas where no or few data po<strong>in</strong>ts reside, creat<strong>in</strong>g mounta<strong>in</strong> ranges for<br />
cluster boundaries. On the other hand, visualized valleys <strong>in</strong>dicate clusters of data s<strong>in</strong>ce small U-<br />
matrix values are observed <strong>in</strong> areas where the data space distances of neurons are small. Distance<br />
based map visualizations (e.g. U-matrix) usually work well for clearly separated clusters; however,<br />
problems may occur with overlapp<strong>in</strong>g clusters. Density-based SOM visualizations display the density<br />
of the data onto the map space via the best-match<strong>in</strong>g neurons. The P-matrix (see Fig. 5(b))<br />
displays the local density measures with Pareto Density Estimation [21]. Neurons with large P-<br />
matrix values are located <strong>in</strong> dense regions of the <strong>in</strong>put vector space and, therefore, areas with<br />
height P-matrix values <strong>in</strong>dicate clusters <strong>in</strong> the data.<br />
3) <strong>Play</strong>er Types: The two map visualizations are complementary and used for cluster<br />
identification with<strong>in</strong> the TRU data. Four ma<strong>in</strong> classes (player types) can be identified as depicted <strong>in</strong><br />
Fig. 5(a) and Fig. 5(b). The best-match<strong>in</strong>g neurons for all 1365 <strong>in</strong>put vector samples are also<br />
represented with small squares of vary<strong>in</strong>g color — different colors correspond to different clusters<br />
of the best match<strong>in</strong>g neurons. On the same basis, Table I presents the number of observations (i.e.<br />
players completed TRU) and percent of neurons belong<strong>in</strong>g to each of the four clusters. Note that<br />
87 (6.37% of the sample) players were not assigned to any cluster s<strong>in</strong>ce their best match<strong>in</strong>g<br />
neurons are placed <strong>in</strong> cluster borders of the ESOM (see Fig. 5).<br />
Fig. 6 illustrates the correspond<strong>in</strong>g component planes of the ESOM. A component plane projects the<br />
relative distribution of one <strong>in</strong>put data vector component (i.e. <strong>in</strong>put vector dimension) to the ESOM.<br />
In the grayscale illustration of those values, white areas represent relatively small values while dark<br />
219