Three papers, and what they mean today
These three publications span the early part of my research: a theory result on interconnection networks, a patent application from my work on Windows developer analytics, and a technical disclosure that, in hindsight, was an early step toward the simulator-driven and synthetic-data methods I care about today. Re-reading them through a modern lens, each one maps onto a question that the field is still working on.
- Hamiltonicity of a General OTIS Network → efficient interconnection and parallel compute
- Filtering Redundant Consumer Transaction Rules → keeping learned models compact
- Optimal simulation of real-world transactions → synthetic data, simulators, and reinforcement learning
1. Hamiltonicity of a General OTIS Network
An OTIS (Optical Transpose Interconnection System) is an optoelectronic parallel computer. Processors are split into groups, linked electronically within a group and optically across groups. The optical links follow a simple transpose rule:
A Hamiltonian cycle is a cycle that visits every node exactly once. It matters because it gives the network fault tolerance and lets it emulate a ring or a linear array, which many parallel algorithms assume.
Prior work handled odd and even n with two separate schemes. Our contribution was a single unified method for both, plus an iterative algorithm:
Why it still matters
The question of how nodes talk efficiently never went away; it just moved up the stack. Today it shows up as GPU and NVLink interconnect fabrics, all-reduce topologies in distributed training, and data-center network design. The same parsimony, do the most with the fewest links, is the design pressure behind every modern supercomputer.
2. Filtering Redundant Consumer Transaction Rules
In the Windows 8 / 8.1 developer analytics work (“Developer Miner”), we mined a data mart of app-sales transactions to surface business patterns to developers, things like which demography, platform, and genre combinations drive purchases. A decision tree over those features produced rule sets, and a real problem emerged: many of the rules were near-duplicates, cluttering the output with the same finding re-expressed dozens of times.
The insight was to remove redundant rules before showing anything to a developer. Formally, each rule covers a set of transactions:
To keep the comparison tractable on large rule sets, rules were bucketed by support level (a “support band”) and clustered with a similarity measure before the subset test:
Why it still matters
Removing redundant learned rules is the same instinct that, decades later, drives weight pruning, network compression, and deduplication of training data. Whether the redundant unit is a mined rule or a near-identical weight matrix, the question is the same: how do you keep only the parts of a learned model that actually carry signal?
3. Optimal simulation of real-world transactions
To test analytic and recommendation systems you need realistic transaction data, but you cannot always touch real users’ records. The disclosure is a defensive publication of a synthetic data generator: it learns business patterns as decision trees (or from a small sample of domain transactions) and then simulates realistic transaction streams from those patterns.
This is, in modern terms, an environment
Here is the connection that makes this disclosure feel current. Generating a stream of events from a model of how a domain behaves is exactly what a reinforcement-learning environment does. An RL problem is formalized as a Markov decision process:
An agent follows a policy, and the environment emits experience as a trajectory:
A simulator produces that trajectory by sampling the transition and reward. The 2011 generator did the same job for analytics, sampling transactions from learned business patterns instead of \(P\) and \(R\). The object is identical; only the notation was different.
“From a small sample of transactions” is model-based RL
The disclosure works even from a small sample of domain transactions. That is the setup of offline and model-based reinforcement learning: learn the dynamics from limited logged data, then generate from the learned model because real interaction is scarce, expensive, or unsafe.
This is the same loop behind world models and model-based RL: a learned simulator stands in for the world so an agent can be trained against it in bulk.
“Without exposing real identities” is the privacy argument for synthetic data
The reason we simulated transactions rather than replay real records was privacy. That exact motive now reappears across modern AI: synthetic datasets for training, simulated environments for robotics (sim-to-real), and generated preference data in RLHF. If you can generate a faithful stream of experience, you never have to touch the real records.
The gap that still matters
The hard part, then and now, is how faithful the simulator is. The learning is only as good as the model of the world underneath it:
In 2011 the world model was a decision tree. Today it might be a neural network trained on video, but the two questions that determined everything back then are still the two that matter: how faithful is the simulator to reality, and how do you guarantee coverage of the rare, important cases? Those are, in a phrase, the same problems that occupy the RL and synthetic-data research I care about now.
The through-line
Read together, the three papers tell one story: compute efficiently, keep learned models compact, and train on data you generate from a model of the world. Those are the same three threads in my work today on efficient language models, model evaluation, and multi-agent systems, now with far better tools than a decision tree.