Abstract
Artificial intelligence (AI) and autonomous technologies have been increasingly embedded into society, so much so that many applications have become common place. The effects these systems can have take place on drastically different scales ranging from shopping recommendations to self-driving cars to military applications. The technical project provides foundational work on the application of AI to the naval combat setting, specifically for location prediction of adversarial ships. Such an application would be a valuable tool to save lives and resources in decision making in naval combat, which is often characterized by swift and tightly coordinated decisions. The science, technology, and society (STS) project attempts to contribute to the growing literature surrounding responsibility attribution in AI and autonomous systems. The lack of clear agency when such systems cause harm intensifies the public mistrust of them and the companies that develop them while limiting or slowing innovation. These projects are loosely coupled as the STS work is not constrained to specific applications of AI however the technical work itself, along with its developers, are subject to analysis under the STS project.
The technical project aims to create a machine learning model to predict the location of adversarial ships in real time based on the known positions of other ships on the battlefield. This project’s purpose is to act as a proof-of-concept to understand and benchmark the capacity of AI when applied to simulated adversarial naval behavior. Two models were constructed, an artificial neural network (ANN) and a Random Forest model, using gameplay data from the online, multiplayer naval combat video game World of Warships. This data was created through a World of Warships tournament where former naval officers acted as commanders of competing fleets in order to simulate the tightly coordinated strategy found in real naval warfare. The models were then used to create predictions that update throughout gameplay that are overlayed on the gameplay map.
Both models proved to be successful in creating predictions of adversary locations, with the ANN slightly outperforming the Random Forest. Specifically, the ANN’s performance was better maintained when assessing the testing Area under the ROC Curve (AUC), despite both models performing comparably on the training and validation data. While neither model performed near perfectly, the performance using only three battles for training indicates these methods have huge potential for future research. To improve performance, more training data should be collected and a richer feature set should be created. The existence of such a tool in naval warfare would provide a huge strategic advantage on the battlefield and hopefully allow the minimization of loss of life and wasted resources.
The importance of understanding attribution in AI and autonomous systems has been intensified by many widely publicized incidents specifically involving deaths as the result of semi-autonomous vehicles. Latour, Callon, and Law’s Actor-Network Theory (ANT) was used to identify and analyze the relevant social groups that exist throughout the life cycle of an AI or autonomous system. An Aristotelian definition of responsible agency was used in the support of ANT. This analysis was shaped through the digestion of current events involving semi-autonomous systems and relevant scholarly discussions centering on agency and ethics in AI and autonomous technology.
The application of ANT spotlighted the highly complex and overlapping nature of possible actors in these systems. Ultimately, the main responsible agent was found to be conditional on the specific circumstances and the interactions between the actors leading up to the incident. The nature of these actors sparked further discussions on current issues in responsibility attribution and distribution including the possible lack of responsible agents, moral crumple zones, and the idea of many hands or things. This discussion did not result in a clear actor but created a framework to understand circumstances and responsibilities that these different actors have that can lead to the transfer of responsibility among them.
The concurrent development of these projects, despite being individual entities and of differing scopes, forces further introspection by each’s contributors. In turn, the creation of the technical work was done under the influence of the consideration of responsibility attribution while the STS work was consistently shaped by the experience of directly working on an AI project.
