Research

My research centers on interpretable AI: designing systems whose decisions can be understood, trusted, and improved by humans. I favor transparent, symbolic approaches over black-box models, and believe that if we can't explain why a system makes a decision, we can't truly trust it. My current interests span program synthesis and symbolic library learning, concept bottleneck models, sample-efficient RL, and general game playing.

Publications

General Game Playing as a Bandit-Arms Problem: A Multiagent Monte-Carlo Solution Exploiting Nash Equilibria

Brandon Mathewe Banda

Oberlin College Honors Papers, 116 (2019)

Paper Code

An interpretable general game player that takes only the rules of a game as input, requires no pre-training, and whose reasoning is traceable to accumulated self-play statistics. Move selection is framed as a multi-armed bandit problem and solved with a two-stage multiagent Monte Carlo search: the first stage prunes losing branches and the second exploits promising ones with a UCB-style policy, converging to Nash equilibria. The player achieves perfect tic-tac-toe, reliably finds winning moves in Nim, and plays strategically in minichess.

  • Rules-only input, no pre-training, and reasoning traceable to accumulated self-play statistics
  • Move selection framed as a multi-armed bandit, solved by a two-stage multiagent Monte Carlo search that prunes losing branches and exploits promising ones with a UCB-style policy
  • Perfect tic-tac-toe play; reliably finds winning moves in Nim; strategic play in minichess

Applied Research

Software Engineer (L4)

Google

  • Led a research agenda to discover techniques for LLMs to dramatically improve open-source developer workflows
  • Proposed unit-test automation as the research objective, designed evaluation frameworks (hybrid embedding + code-coverage evals) and experiments, and developed a graph-structured reasoning framework for automated test generation across AndroidOS
  • Built structured, hierarchical datasets of Android's source code to improve one-shot performance and fine-tuning
  • Collaborated with DeepMind and CoreML teams to define data specifications and quality criteria for training Google's internal models on open-source code—the first time these models were trained on open-source codebases
  • Developed an unsupervised root-cause analysis method for Android OS infrastructure errors using T5X embeddings and density-based clustering, reducing incident response time by 90%

Research Interests

Program Synthesis & Symbolic Regression

Symbolic, non-neural library learning that accumulates data-efficient, reusable abstractions across domains. Minimum description length and e-graphs to induce concise generalized solutions, treating compression as learning.

Concept Bottleneck Models

Using program synthesis as the concept generator for concept bottleneck models in place of hand-picked or LLM-generated concepts, and testing what that changes for predictive accuracy and interpretability across problem spaces.

Reinforcement Learning & General Game Playing

Sample-efficient methods that make symbolic approaches tractable for intractably deep search problems.