1 Climb a-GAN: Generation of Rock Climbing Problems 59 2 60 3 ∗ 61 4 Cove Geary Joseph Valdez 62 5 [email protected] [email protected] 63 6 Yale University Yale University 64 7 New Haven, Connecticut New Haven, Connecticut 65 8 66 9 67 10 68 11 69 12 70 13 71 14 72 15 73 16 74 17 75 18 76 19 77 20 78 21 79 22 80 23 81 24 Figure 1: Example of a climber using a Moonboard 82 25 83 26 ABSTRACT ACM Reference Format: 84 27 Many rock climbing gyms have adopted a new type of rock climbing Cove Geary and Joseph Valdez. 2019. Climb a-GAN: Generation of Rock 85 Climbing Problems. In Proceedings of ACM Conference (Conference’17). ACM, 28 wall: the MoonBoard. The MoonBoard is a standardized interactive 86 New York, NY, USA, 6 pages. https://doi.org/10.1145/0000000.0000000 29 training wall of 11 x 18 holds that are identical in each gym. The 87 30 MoonBoard website provides a space for users to create new custom 88 31 paths, upload them, and even rate other paths. Using an ACGAN, it 89 32 is possible for a user to be able to input a difficulty for a desired rock 1 INTRODUCTION 90 33 climbing path and have it generate a new suggested path according Our project is motivated by two challenges in indoor rock climbing: 91 34 to that difficulty that is compatible with the MoonBoard. Current generating new climbing paths (route- or problem-setting); and 92 35 GitHub projects for MoonBoards generate a random path and then understanding what makes a particular problem challenging and 93 36 classify the difficulty, so this would be approaching it in reverse. fun. 94 37 By generating a path based on a given difficulty, it will be easier for Although the full space of possible mediums (wall size, arrange- 95 38 rock climbing users to find a path they would be able to attempt ment, etc.) for a climbing problem is huge, perhaps the best candi- 96 39 (and enjoy). The classifiers in the ACGAN are the different levels date medium for starting to approach this algorithmically would be 97 40 of difficulties of the rock climbing routes.Not The ACGAN for will output distribution.the MoonBoard. The MoonBoard is a standardized bouldering wall 98 41 an image of a MoonBoard withUnpublished the starting holds circled in green, workinglocated in gyms throughout draft. the world, such that every MoonBoard 99 42 the intermediate holds circled in blue, and the ending holds circled has the same features for hands and feet (holds) in the same place, 100 43 in red. arranged in a 11x18 grid. Climbing paths can be customized and 101 44 created by users on an online community. The MoonBoard website 102 45 KEYWORDS hosts an online community where users can create problems, as- 103 46 datasets, CGAN, conditional GAN, path generation sign them a difficulty, and share them publicly for others to view.A 104 47 problem on the MoonBoard is defined as a sequence of holds along 105 48 ∗Both authors contributed equally to this research. the grid. Each problem is given a grade (difficulty level) and a rating 106 49 (0-3 stars). There is even an app where users can upload paths so 107 50 Permission to make digital or hard copies of all or part of this work for personal or 108 classroomUnpublished use is granted working without fee draft. provided Not that copies for distribution.are not made or distributed there exists a healthy amount (many thousands) of accessible data! 51 for profit or commercial advantage and that copies bear this notice and the full citation However, since users currently have to manually design paths 109 52 on the first page. Copyrights for components of this work owned by others than ACM and rate their difficulty, we were hoping this process could instead 110 must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, 53 111 to post on servers or to redistribute to lists, requires prior specific permission and/or a be streamlined with the aid of technology. It would be incredibly 54 fee. Request permissions from [email protected]. helpful if instead of creating your own path or even searching 112 55 Conference’17, July 2017, Washington, DC, USA through thousands of paths to find one that matches your profi- 113 © 2019 Association for Computing Machinery. 56 114 ACM ISBN 978-x-xxxx-xxxx-x/YY/MM...$15.00 ciency with bouldering, a user could simply input their desired 57 https://doi.org/10.1145/0000000.0000000 difficulty and have a path automatically generated for them. 115 58 2019-12-11 22:52. Page 1 of 1–6. 116 Conference’17, July 2017, Washington, DC, USA Geary and Valdez 117 This might be able to be achieved through a GAN, or Generative classifying MoonBoard problems [1]. Comparing Naive Bayes, Soft- 175 118 Adversarial Network. GANs are given training data, then using two max Regression, and CNN (with ordinal output), they found the 176 119 networks working against each other, eventually learns to generate CNN to be the best-performing. 177 120 new data similar to that of the training set. The two neural networks 178 121 are referred to as the generator and the discriminator. The job of 1.2 Related Work: ACGANs 179 122 the generator is to learn to generate new data capable of fooling Taking this further, it would be helpful for users to be able to 180 123 the discriminator into thinking it was from the initial training generate a path according to their desired difficulty, making it 181 124 data. The job of the discriminator is to be able to determine where easier to find a path they will enjoy. This could be possible byusing 182 125 the generated data is "real" or "fake", or whether the data is from an ACGAN, shown in the following figure. 183 126 the initial training data or fabricated by the generator. After many 184 127 iterations of this training, the generated data should so similar to the 185 128 training data that it fools the discriminator. The model architecture 186 129 is shown in the figure below. 187 130 188 131 189 132 190 133 191 134 192 135 193 136 194 137 195 138 196 139 197 140 198 141 199 142 200 Figure 2: Model of a GAN from 143 201 https://pathmind.com/wiki/generative-adversarial- 144 202 network-gan 145 203 146 204 147 205 Using this method, we could train a GAN using the thousands 148 206 of user-created MoonBoard problems with the intention of being 149 207 able to generate new problems similar (and equally traversable) to 150 208 the problems in the online community. 151 209 152 210 153 1.1 Related Work: Moonboards 211 154 As far as we can tell, very little work has been done thus far in the 212 155 space of climbing problem/route generation. Further, anecdotally 213 156 speaking, the task of route-setting tends to be seenNot by climbers for as distribution. 214 157 one that requires more than justUnpublished practical skill—it is one that draws working draft. 215 158 on creativity. For this reason, MoonBoard problem generation poses 216 159 an exciting challenge for generative models. 217 160 As mentioned, the space of climbing problem generation, and 218 161 even classification, seems to be relatively untouched. Some smaller 219 162 GitHub-hosted projects seem to apply neural networks to classify 220 163 the difficulty of existing MoonBoard problems (e.g. with anMLP, 221 164 with a CNN, with an LSTM). We found one one GitHub-hosted 222 165 project that seems to have experimented with generating novel 223 166 MoonBoard problems. However, each of these repositories appears 224 167 minimally-documented, and the results of each project is unclear. Figure 3: Model of an ACGAN 225 168 For the one generative project, problem generation and difficulty 226 169 classification are done in two separate steps, so the starting “hold” An AC-GAN, or Auxiliary Classifier Generative Adversarial Net- 227 170 is chosen at random to begin problem generation, and problem work, is an extension of the classic GAN method. In an ACGAN, 228 171 difficulty is assessed post-generation. the data is also given a classifier or label. The discriminator then 229 172 In one paper from an undergraduate machine learning final has one more job in addition to being able to determine whether 230 173 project, the authors present a more methodological approach to data is "real" or "fake". The discriminator now must also be able 231 174 2019-12-11 22:52. Page 2 of 1–6. 232 Climb a-GAN: Generation of Rock Climbing Problems Conference’17, July 2017, Washington, DC, USA 233 to predict the class label of the given data. This then allows the 291 234 generation of data of a specific type, according to the class label. 292 235 In the case of rock climbing problems, the different difficulties 293 236 of the climbing problems could be used as the different classes. 294 237 The ACGAN could then generate paths similar to those of the 295 238 same difficulty in the training data set, which in this case would 296 239 be the collection of user-made paths on the MoonBoard website. 297 240 The generated image will then be a generic image of the 11 x 18 298 241 MoonBoard grid, and the holds that are used in the path will be 299 242 circled.