My research focuses on analyzing and imitating player behavior in multiplayer, first-person shooter (FPS) video games like Counter-Strike: Global Offensive (CSGO). I document my research successes and challenges on my blog. Through this process, I've honed in on the key idea that imitating human behavior enables analyzing human behavior.
- Problem Selection - how I identified analyzing player behavior in CSGO as a good problem that is both important, as it enables understanding complex behaviors driven by expert knowledge, and feasible, as there's ample data.
- Initial Approaches - how I determined that new solutions are necessary, as combining standard graphics and analytics techniques is insufficient to understand player behavior. I tried to use these techniques to identify cheaters using their reaction times. I computed accurate reaction times by rerendering game traces with custom shaders, addressed confounding technical factors by modifying the renderings to account for network latency, and featurized confounding behavior factors by using ray tracing to identify where players may predictively aim. While these techniques are precise and interpretable, they fail to provide satisfactory results as they only look at a small subset of game state. Human behavior is situationally dependent, so an individual player's decisions must be analyzed in the context of all players' game state.
- Technical Challenges - how I addressed technical challenges that arose during my journey, like extracting weapon recoil state and player head position from the game engine.
-
Progress (and Eternal Glory) With New Approaches - how I'm
solving the problem. Rather than directly analyzing human behavior, I attack the problem
indirectly.
- In some situations, "indirectly" means forcing the players to perform their own analyses. I collaborated with Activision on hallucinations, an anti-cheat technique that tricks cheaters into self-identifying by reacting to fake enemies that aren't visible to legitimate players. Cheaters can only avoid detection by analyzing which players in the game state are real and which are fake. I presented hallucinations at GDC 2023.
- In other situations, "indirectly" means focusing on an intermediate problem: imitation rather analysis. Once you can imitate human behavior, then you can analyze the imitations and use them in novel applications like coaching. I've created a hand-crafted bot as a structure for imitation learning and attempted to learn an aim controller using data driven techniques.
I received a NSF Graduate Research Fellowship and a Stanford Graduate Fellowship in Science and Engineering. I will be graduating and looking for a full-time job this year.
Other Research
Designing efficient, application-specialized hardware accelerators requires assessing trade-offs between a hardware module's performance and resource requirements. To facilitate hardware design space exploration, we describe Aetherling, a system for automatically compiling data-parallel programs into statically scheduled, streaming hardware circuits. Aetherling contributes a space- and time-aware intermediate language featuring data-parallel operators that represent parallel or sequential hardware modules, and sequence data types that encode a module's throughput by specifying when sequence elements are produced or consumed. As a result, well-typed operator composition in the space-time language corresponds to connecting hardware modules via statically scheduled, streaming interfaces.
We provide rules for transforming programs written in a standard data-parallel language (that carries no information about hardware implementation) into equivalent space-time language programs. We then provide a scheduling algorithm that searches over the space of transformations to quickly generate area-efficient hardware designs that achieve a programmer-specified throughput. Using benchmarks from the image processing domain, we demonstrate that Aetherling enables rapid exploration of hardware designs with different throughput and area characteristics, and yields results that require 1.8-7.9x fewer FPGA slices than those of prior hardware generation systems.