Date of Award
Santa Clara : Santa Clara University, 2022.
Computer Science and Engineering
David C. Anastasiu
Today, artificial intelligence is used to solve problems as varied as driving cars to diagnosing diseases. However, it still has major pitfalls. The most common form of AI today, machine learning, can produce near-certain predictions for some tasks, yet completely fail at a related but different task. Additionally, such algorithms cannot do reasoning for complex problems, are not well explainable, require massive training datasets, and are sensitive to new data.
One solution to these problems is combining today’s cutting-edge neural networks with an older idea in AI: symbolic AI. Such so-called “neuro-symbolic AI” systems can be understood at a high level as using a neural network as the “sensory cortex” and a question answering system as the “prefrontal cortex” to do reasoning. Neurosymbolic AI (NSAI) has already been shown to excel at visual question answering (VQA), a task that involves answering a question about something happening in an image. Historically, neural networks have been particularly weak in this area. Researchers are just beginning to explore the power of neuro-symbolic AI in approaching video VQA to learn physical dynamics. In the future, such algorithms may play a role in making more flexible and intelligent AI, such as for autonomous driving systems.
Right now, neuro-symbolic AI has many areas for expansion across a wide range of applications. This paper will seek to apply NSAI to strategic gameplay where agents have the same information a human player has—just an image. To enable testing abstraction and high-level reasoning, I develop the Blokboi game as an AI training and testing tool. Blokboi pushes an AI agent to learn relationships that strain machine learning, and provide an environment in which the AI can be tested for their ability to learn compound interpretation and reasoning. This research expands the applications of NSAI, bringing hybrid artificial intelligence increasingly close to real-world scenarios.
Clay, Dorian, "Expanding Neuro-Symbolic Artificial Intelligence for Strategic Learning" (2022). Computer Science and Engineering Senior Theses. 228.