SMC/particle filtering in life

[hyunjimoon]

My approach of seeking advice from Bob, Andrew, Mitzi, Mike, Paul, Martin, Charles, Ben, Simon, and other Stan family members is intentional. It's based on a relatively informative covariance metric I've developed for these nine individuals in a desire-belief parameter space.

Drawing an analogy to SMC methods:

1. Particle representation: Each of the nine individuals (Bob, Andrew, Mitzi, etc.) represents a "particle" in the space of Stan development and probabilistic programming philosophy.

2. Importance sampling: Your insights as a historian of Stan provide importance weights, helping to refine the posterior distribution of Stan's evolution.

3. Hierarchical model: My mind has constructed a hierarchical model of Stan as an organism:

   • Hyperparameter: Stan culture
   • Parameters: Each actor's strengths and philosophies

4. Sequential updating: Each observation or piece of advice from the nine "particles" updates the model sequentially, similar to how SMC methods update particle weights over time.

5. Resampling: By synthesizing insights from diverse perspectives (e.g., Ben's hierarchical Bayes with quantile priors, Mike's HMC diagnostics, Paul's simulation-based inference, Simon's SMC work), I'm effectively resampling to maintain a diverse and representative set of particles.

6. Proposal distribution: My questions and prompts serve as proposal distributions, guiding the exploration of the parameter space of Stan's future development.

7. Intermediate target distributions: The evolving understanding of Stan's trajectory can be seen as a series of intermediate target distributions, gradually approaching the true posterior of Stan's future state.

This SMC-inspired approach allows me to trace weight updates for each individual's contributions and average them, potentially providing a better estimate of the future of probabilistic programming – an area where I hope to contribute.

Regarding your question about what's "dynamic," my advisor's motto, "All advantage is temporary," encapsulates this concept well. It speaks to the constantly evolving nature of the field and the need for adaptive strategies.

The goal of Stan talk was indeed to present options for graceful evolution (or disintegration, if necessary) which Stan, like any long-lived project, may face at some point. This approach acknowledges the dynamic nature of the field while seeking to maximize the longevity and impact of Stan's contributions.