IIIT Work Summary May 2024

The following is an incomplete summary of work I've done while in the Dual Degree (B.Tech Computer Science, MS by Research in Exact Humanities) course at IIIT-Hyderabad, from 2011 to present. I have focused largely on my MS work for the sake of brevity. My transcript and IMS submissions to date will cover any omissions.

I have not made any publications in outside venues or earned any awards while in my course of study. I was invited to attend a short workshop called CFAR for Machine Learning Researchers conducted by the Center for Applied Rationality in the autumn of 2016 to discuss my work in context of the value complexity problem in AI alignment research.

Most of the utility of my course of study in my present career as a site reliability engineer has come from taking the Operating Systems course, and self-study that resulted from it. I also received valuable, if less tangible, benefits from the portions of my course that were concerned with philosophy, history, and society.

I chose my work in the latter years of my course out of interest in the fields of game theory and logic, with the aim of exploring a point of confluence between computer science and philosophy. Specifically, I was connecting the philosophical idea of heterotopias, first put forward by Foucault, which discusses the dynamic nature of systems that interface heavily with society, with the field of game semantics, which deals in formal representations of strategic action.

I hope to use the skills and knowledge I've built in my research work in my career in the next five years. I see several possibilities for this: notably, using modal and temporal logics to study how to better design and administer large-scale computing systems; and applying of a game-theoretic understanding of semiotics to identify dynamic properties of large codebases.

Coursework

• Exact Humanities Course Syllabus, 2011-2016 Please find an abridged reading list of the material referenced here.

TAships

• Principles of Programming Languages, 2018.

• Site and Art, 2015.

Projects

• Weed-grower, 2013 For the Game Design course, I designed a mobile game based on Conway's Game of Life.

• Narrative structures and metonymy, 2014 I wrote a script for a short film titled Against the Ground. This script explores how metonymy interacts with medium by making the tools used to manipulate 3D objects in 3D modeling and animation software a part of the film setting. The work is a theoretical investigation into how affordance becomes narrative - how a setting is part of what constitutes a plot, and how tools and environment dictate what can happen, and therefore cannot be analytically disentangled from what does happen.

• Translation of Chapter 1 of Dasarupaka, 2015 For my final project in Classical Text Reading I, I translated a portion of the Dasarupaka by Dhananjaya.

Self study courses

• Modal Logic

• Graph Grammars

• Analysis of postmodern philosophers: Hannah Arendt and G.E.M. Anscombe

Research

Prof. PRK Rao originally suggested to me the research topic of heterotopia, and offered The Heterotopia of Facebook as a starting point. I began by approaching this problem in terms of the semiotics of Facebook - how the system encodes social actions (e.g. the Like button), and how signals like this get exapted for other purposes and undergo semantic drift. I wanted to represent this type of effect formally, and to modify the classical representation of games in game theory to allow for player-driven changes to the game specifications. I was also motivated by Chen & Micali, which discusses motivated mechanism design, and Ausubel…, which shows that the Vickrey auction is vulnerable to cheating by collusion, and expresses the “cheat” as a minor edit to the game specification. This in turn suggests a notion of edit distance across game mechanisms, and from there the possibility of constructing a stability metric in the induced space. Put together, the thesis was: mechanism changes can be strategic behaviour, so there ought to be a way to represent that behaviour as part of the game strategy. Depending on how one operationalizes this project, several of the hard problems in game theory are subproblems to this one:

• Complexity limits on the size of a game, imposed by the backwards induction solution concept

• How to define dominant strategies, and consequently equilibria, for a long-running game

• The fragility of most existing solution concepts to small changes in the game specification

I came up with several applications and approaches to this problem between 2017 and 2022.

• Metagames: My first approach was to create a game where players could choose among games in a restricted way: exploring choices in a finite number of finite games, or an iterated choice of finite games. This reduces to an alternate way of representing iterated games, and can't be fairly described as having "dynamics" or "player edits" without further characterization. This approach runs fairly quickly into the problem of Hypergame paradox, popularized by William Zwicker in 1987: how to ensure that a self-referential structure of this kind is restricted enough to be finite, or remain within computability bounds that allow it to be useful?

• ATL: Alur… put forward Alternating-Time Temporal Logic (ATL) as a way to express games which can be expressed as a concurrent game system - in other words, a graph that functions as the generator for an extensional game, with a coalition-aware solution concept that can be expressed and deduced on the graph and proven to hold for the generated game tree. I wrote a representation of ATL in Coq theorem prover, and used it to express a dominance relation over the set of ATL formulae that described the graph. Unfortunately, this relation was not computable in its original form.

• Nomic: a game first specified in Suber,as an illustration that a self-modifying ruleset cannot be written to preclude states where the rules become uninterpretable. Nomic is specified as an InitialSet of rules written in English and meant to be played by human agents; in traditional conceptions of homo rationalis there is no reason to play this game and no way to write a dominant strategy (as the InitialSet puts winning purely to chance). I attempted to create a model of just one element of the Nomic ruleset: players can both read and write the game specification; and provide an example game of Nomic (with a more restrictive InitialSet) for which a winning strategy is computable. Significant prior art exists in formally modeling Nomic, but none built to verify the thesis of the original paper: does self-modification inevitably lead to un-interpretability?

I didn't succeed in writing the formal representation, having run aground on the difficulties of finding a fair, sufficiently general language for rules and edits.

• Preference meronomy: or part-whole relationships in coalitional strategies. Can a coalition be considered to have a preference? Since it possesses a winning strategy, and can be reasoned about in counterfactuals, one can induce a preference relation over outcomes that corresponds to a coalitional strategy. One can therefore explore how individual preferences compose into collective preferences.

• Property-based preferences: using an intermediate representation to represent player reasoning about outcomes. Dietrich & List describe a way to construct this, which I spent some time trying to use to represent player reasoning: by having the "properties" be true claims about the reachability of various outcomes from the node in question.