Cost function
CARDAMOM is a Bayesian modeling framework designed to compute the posterior distribution of model parameters, based on Bayes’ Theorem:
In this context, \(P(\mathbf{x} | \mathbf{O})\) represents the posterior distribution of the parameter vector \(\mathbf{x}\) given the observation matrix \(\mathbf{O}\). CARDAMOM approximates \(P(\mathbf{x} | \mathbf{O})\) using a Markov Chain Monte Carlo (MCMC) sampler. Here, \(P(\mathbf{x})\) denotes the prior distribution of the parameters, while \(P(\mathbf{O} | \mathbf{x})\) is the likelihood of observing \(\mathbf{O}\) given the parameter values in \(\mathbf{x}\).
In CARDAMOM, \(P(\mathbf{x})\) is expanded into two components: the parameter prior range and the ecological dynamic constraints (EDC), resulting in the following expression:
To simplify the cost function, we take the logarithm of the posterior distribution, yielding:
This logarithmic transformation converts multiplicative relationships into additive ones, thereby reducing the complexity of the MCMC sampling process by narrowing the search space.
The likelihood function \(P(\mathbf{O} | \mathbf{x})\) follows a log-normal distribution and is expressed as:
Where \(M\) represents the model predictions, \(O\) represents the observations, and \(\sigma\) denotes the uncertainty associated with the observations.
The prior distribution \(P_{\text{prior}}(\mathbf{x})\) is modeled as a log-uniform distribution.