Learning constructive primitives for procedural content generation
Peizhi Shi · Research Explorer (The University of Manchester) · 2019
Introduction 4.4 Low-quality game segments.(A) Unplayable segments.(B) Segments with unbalanced resources.(C) Segments with unexpected difficulty spikes.(D) Segments with unreachable resources.(E) Aesthetically unappealing segments. . . . . . . . . . . . . . . . . .103 4.5 An example of how to generate a complete game level via CPs. . .106 4.6 Exemplary game segments in different clusters. . . . . . . . . . . .110 4.7 HTER on validation sets achieved by different classifiers during active learning. . . . . . . . . . . . . . . . . . . . . . . . . . . . .114 4.8 Error rates on validation sets achieved by different classifiers.(A) WRFs, (B) SVMs, (C) ANNs, (D) DT . . . . . . . . . . . . . . .116 4.9 Error rates on validation sets achieved by different classifiers with different batch sizes.(A) WRFs, (B) SVMs, (C) ANNs, (D) DT .117 4.10 Error rates on validation sets achieved by different query strategies during active learning. . . . . . . . . . . . . . . . . . . . . . . . .118 4.11 Error rates on validation sets achieved by different validation set construction approaches. . . . . . . . . . . . . . . . . . . . . . . .119 4.12 Test game segments labelled by our classifier.(A) Correctly classified positive segments.(B) Segments leading to false negative error.(C) Correctly classified negative segments.(D)Segments leading to false positive error. . . . . . . . . . . . . . . . . . . . .120 4.13 Quality assessment results with different metrics for SMB level generators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .125 4.14 The distribution of the number of low quality game segments in a procedural level. . . . . . . . . . . . . . . . . . . . . . . . . . . . .126 4.15 Expressive ranges of our level generator corresponding to different controllable parameter values. . . . . . . . . . . . . . . . . . . . .127 4.16 Exemplar levels generated with different linearity values.(A) linearity = 1.(B) linearity = 2. (C) linearity = 3. . . . . . . . . . .128 4.17 Exemplar levels generated with different leniency values.(A) leniency = 1.(B) leniency = 2. (C) leniency = 3. . . . . . . . . . .128 4.18 Exemplar levels generated with different density values.(A) density = 1.(B) density = 2. (C) density = 3. . . . . . . . . . . . .128 4.19 Statistics of completion rates achieved by 14 agents on adaptive and static game sets. . . . . . . . . . . . . . . . . . . . . . . . . .131 9 4.20 Completion rates of different agents: adaptive (θ opt = 0.98) vs. static.(A) Oliveira's agent.(B) Lopez's agent.(C) Baumgarten's agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .131 4.21 Completion rates of different agents: adaptive (θ opt = 0.7) vs. static.(A) Oliveira's agent.(B) Lopez's agent.(C) Baumgarten's agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .132 4.22 Exemplar procedural levels generated by our CP-based DDA algorithm for different agents.(A) Baumgarten's agent.(B) Lopez's agent.(C) Oliveira's agent. . . . . . . . . . . . . . . . . . . . . .133 5.1 Game segments in CUBE 2. . . . . . . . . . . . . . . . . . . . . .136 5.2 Game design elements in Cube 2. (A) Powerups (weapons and health pack).(B) Enemies.(C) Doors.(D) Rooms.(E) Entrance and exit.(F) Box.(G) Blocks.(H) Stairs.(I) Walls. . . . . . . .137 5.3 A game segment and its representation. . . . . . . . . . . . . . . .138 5.4 Patterns for wall-block.(A) Pattern 1. (B) Pattern 2. . . . . . . .140 5.5 Low-quality game segments.(A) Unplayable segments.(B) Segments with unbalanced resources.(C) Segments with unexplainable difficulty spikes.(D) Segments with unreachable resources.(E) Aesthetically unappealing segments. . . . . . . . . . . . . . .142 5.6 Game map generation process.(A) A map graph.(B) A divided graph.(C) A divided graph and corridor locations.(D) A map without corridors.(E) A map blueprint.(F) A final map. . . . . .145 5.7 Exemplary game segments in different clusters. . . . . . . . . . . .150 5.8 HTER on validation sets achieved by different classifiers during active learning. . . . . . . . . . . . . . . . . . . . . . . . . . . . .151 5.9 Error rates on validation sets achieved by different classifiers.(A) WRFs, (B) SVMs, (C) ANNs, (D) DT . . . . . . . . . . . . . . .153 5.10 Error rates on validation sets achieved by different classifiers with different batch sizes.(A) WRFs, (B) SVMs, (C) ANNs, (D) DT .154 5.11 Error rates on validation sets achieved by different query strategies during active learning. . . . . . . . . . . . . . . . . . . . . . . . .154 5.12 Error rates on validation sets achieved by different validation set construction approaches. . . . . . . . . . . . . . . . . . . . . . . .156 10 5.13 Test game segments labelled by our classifier.(A) Correctly classified positive segments.(B) Segments leading to false negative error.(C) Correctly classified negative segments.(D)Segments leading to false positive error. . . . . . . . . . . . . . . . . . . . .