In this work, we created fast and slow thinking agents for arbitrarily complex action spaces by innovating deep learning and search algorithms. The agents Illustrated an ability to learn design strategies from scratch and effective zero-shot generalization over multiple unseen problems.

In this work, I developed an AI recommender tool that guides human decision-making in design. This study evaluates the temporal aspect of human designers’ confidence in AI. The work demonstrates that designers have a fickle but considerably high level of confidence in AI potentially leading to suboptimal decisions.

This work evaluated the perception of humans on AI teammates by masking some AI as humans while measuring hybrid team dynamics. The analysis demonstrated average designers over-estimate AI’s proficiency while high-performing designers trusted humans more.

Developed a hierarchical, order-invariant, and sampling architecture to represent policies in arbitrarily complex action spaces. Decomposed policy representation into sampling feasible actions and generating probability values over state-action pairs. Illustrated a significant improvement (53% more) in predicting exact human actions using images over Deep Q-Networks.

Developed a framework that integrates data-driven visual intuition (as exploitation) with search (as exploration) for design. Illustrated a methodology to augment design decision-making on unseen problems by introducing a goal-directed search. Demonstrated 200% better performance than average humans in both original and an unseen variant of truss design problem.

Developed a deep learning-based real-time AI guidance tool that is trained to imitate average humans in a design study. Demonstrated cognitive over-load and ineffective human-machine interfacing as limitations of real-time collaborative design.

Enabled a novel learning from pixels approach using convolutional networks to capture design generation strategies from data. Developed a framework that learns to predict future design states by imitating human decision-making without feedback info. Demonstrated an ability to learn arbitrary goals from sequential data as agents generate human-level designs without feedback.

Developed a probabilistic model to represent action sequencing relationships capturing diverse skilled decision-making behavior. Demonstrated effective transfer of diverse action-sequencing strategies across design problems showing characteristic behavior.

Developed a data-driven extended pattern search optimization algorithm to determine the optimal configuration of aerial vehicles.