A Supervised-Reinforcement Learning Model for Automated Clash Resolution in the Construction Industry

Clash Coordination is a crucial step in ensuring timely and cost-effective project delivery. While software tools like Navisworks and Solibri have improved the process of aggregating models and conducting clash tests and categorization, resolving clashes remains a slow and manual task. The reason for this slow process can be attributed to the meticulous nature of the process where design coordinators need to ensure that resolving one clash does not lead to new clashes. With the advent of machine learning and its application in construction, more research is being conducted to automate construction tasks to increase productivity and reduce the cost of the project. One such task currently being researched is to automate clash resolution. Researchers have explored the use of machine learning, specifically supervised learning, to automate clash resolution with successful outcomes. A search of the Web of Science database shows 7 publications that discuss the automation of clash resolution, automation of clash correction sequence, and automation of selection of relevant clashes. The authors to further analyze the content of these publications used VOSviewer to create a word map of keywords contained in the title, keywords, and abstract fields of these publications to analyze word co-occurrence. The word co-occurrence analysis revealed that the publications have explored supervised learning as the machine learning category of choice for automating clash resolution. However, the same analysis also showed the lack of terms such as data scrubbing, feature selection, feature engineering, and domain knowledge. These terms are an essential part of developing a machine-learning model. This analysis led the authors to believe that even though research is being conducted to automate clash resolution, a systematic approach to develop a machine learning model to support the automation of clash resolution is missing. Also, though these researches show significant accuracy in terms of automating clash resolution, they fail to justify the selection of their feature and label space. Another limitation of the current state of the art is that the effectiveness of supervised learning in automating tasks is limited by the availability of a large amount of labeled data, often unavailable. To address these research gaps, in this dissertation the author's first contribution to the body of knowledge is a phased systematic approach to develop an automation model for clash resolution. Since in machine learning selection of appropriate feature and label space is critical in developing an optimum and explainable solution, it is crucial to identify features that accurately represent a clash and also represent the factors industry experts consider when resolving the clash. Along with features, labels need to be selected as well to represent clash resolution options available to the industry. To achieve this in chapter 2 the author using modified Delphi captured the domain knowledge that industry experts utilize to resolve clashes. Factors considered by industry experts to decide on how clashes are resolved and options to resolve clashes are extracted from the domain knowledge. As a result of this research, the author identified 23 factors that industry experts consider when resolving clashes and 5 options available to resolve the clash. The work concludes by identifying factors and options that can serve as features and labels for a machine-learning algorithm to automate clash resolution. Once features and labels are identified the author in chapter 3 discusses the development of a prediction model to predict clash resolution options for a given clash. The discussion is focused on individual steps involved in the creation of machine learning models like data collection, data pre-processing, and machine learning algorithm development and selection. The author also addresses common challenges in the development of machine learning models like class imbalance and availability of limited data. The author utilizes a multi-label synthetic oversampling method (MLSOL) to generate different percentages of synthetic data to account for class imbalance and limited datasets. Using this dataset, the author then trained five different supervised learning algorithms and reported their accuracy. Based on this work the author concluded that increasing the dataset with 20% of synthetic data and using an artificial neural network to develop the machine learning model to automate clash resolution generated the best result with an average accuracy of around 80%. To address the limitation of using only supervised learning and as a second contribution to the body of knowledge, the author in chapter 4 proposes the use of reinforcement learning to train a Deep Q Network (DQN) agent capable of learning how to resolve clashes through interactions with a Building Information Model (BIM) environment containing clashes. The work discusses the implementation of a dynamic reward function to guide the agent in making decisions based on industry best practices. Additionally, it outlines the setup of the interaction between the agent and the environment to facilitate learning. Considering that reinforcement learning requires a significant amount of time to develop knowledge, the author also tested the effect of using a pre-trained supervised learning model to initialize the reinforcement learning policy function and guide knowledge exploration. This approach resulted in three variations of supervised-reinforcement learning. The supervised learning model used in this research demonstrated an accuracy of 31%. To demonstrate the utility of reinforcement learning in training an agent, the authors plotted graphs showing the number of clashes resolved per episode and the cumulative reward received per episode. The clashes resolved by the agent in this research were limited to clashes between ducts and pipes. These graphs illustrated that with each successive episode, the agent became increasingly proficient at resolving clashes. Among the variations of supervised-reinforcement learning, the one that exhibited the most stable learning graph utilized the weights of the supervised learning model to initialize the policy function of reinforcement learning. This research confirmed that reinforcement learning can be employed to train an automated model instead of relying solely on supervised learning, especially in scenarios where limited or no clash resolution data is available. Moreover, pre-training reinforcement learning using a supervised learning model led to more consistent learning outcomes. The research presented in this dissertation focuses on the holistic development of a machine learning model to automate clash resolution. By identifying appropriate features and labels before training the model the author ensures that the automation model accurately captures industry best practices and is explainable. Furthermore, by utilizing a systematic approach towards the development of a machine learning model the author addresses common challenges in developing a machine learning model and how we can overcome them. Lastly, through the utilization of supervised reinforcement learning the author proposes an alternative learning algorithm that can learn how to resolve clashes with fewer labeled examples through Building Information Model (BIM) interaction and with a more steady learning rate than reinforcement learning alone.