Home

model

prior choice

• [uninformative priors] (https://en.wikipedia.org/wiki/Prior_probability#Uninformative_priors)
  • maximal entropy (other objectives?) given the conditions we know:
    • mle estimate -> given mode
    • positivity
    • maybe desired variance
  • => gamma distribution?

claudias model

• wrong hill parameters?
• wrong rejection step

effect of rho on exploration/exploitation

• larger rho leads to higher gradients far away from peaks
  • => high acceptance rate
  • connection to simulated annealing?
• iteratively reduce rho?

mamba

• cache result of lpdf

mcmc

nonsymmetric (lognormal) proposal

• results from transforming constrained to unconstrained sample space
• median (0), mode (0), mean (?) -> implications for mcmc?

mcmc

• why q~N(x,I) => N steps take us sqrt(N)?
• connection to stochastic gradient?
• sparse grid as proposal?
• simulated annealing vs stochastic tunneling vs lpdf cutoff (prohibit low prob areas)
• genetic mcmc?