Many ecological processes are best represented by individual-based models (IBMs), where processes at the scale of individual scale of organisms and their actions are modeled explicitly. Such processes can pose a problem for management or control; our observational scale, control levers, or management goals may be limited to coarser scales than the individual.

IBMs suffer from the curse of dimensionality. The number of possible states of the system, which consists of the possible states and combinations thereof of all individuals, is enormous. This makes it difficult to apply control techniques or various levels of system analysis.

One approach to this problem is to apply optimal control techniques to an aggregrated model (AM) that represents an IBM at the population with some loss of accuracy [@Federico2013]. These techniques require an AM that adequately approximates the IBM, though. In particular, it can fail in cases of structured, heterogenous populations of individuals.

@Kevrekidis2009a describe a set of methods for analyzing IBM-like systems at scales coarser than the individual. In this approach, one contructs a “coarse time-stepper” that simulates the evolution of the system through time at the population level. This stepper consists of an IBM, a “lifting” function, and a “restriction” function. The functions translate between coarse and individual scales. The “restriction” function summarizes the distribution of individuals in small set of moments. The “lifting” function generates distributions of individuals from these moments. Short bursts of the IBM simulate the system in between. The time stepper can be used to estimate derivatives of the coarse-level system. Where IBMs are stochastic, this can be accomplished by repeated lifting and bursts of the IBM to obtain mean values.

This approach, known as “equation-free” or “multi-scale” modeling, allows simulation at coarse scale by forward integration of the estimated derivatives rather than full simulation of the IBM. The coarse time stepper can also be used for other numeric analyses, such as mapping equilibria and bifurcation points of the system. Equation-free modeling was developed for simulation of systems in physical chemistry, where molecular-level interactions had to be simulated at the “beaker” level. In the ecological context, it has been applied to swarm dynamics [@Raghib2010a] and the evolution of strains of disease [@Cisternas2004]. The technique hinges on careful selection of the coarse-scale variables, and a separation of time scales between these variables and higher moments of the distributions of individuals, which must come into equilibrium under short bursts of the IBM.

Here I propose an application of equation-free techniques to optimal control in an ecological context.

Spatial point processes

@Brown2004 describe a spatial point process which represents the dynamics of a disease spread through a plant population. Plants are represented as discrete points in continuous space, and each individuals’ infection and mortality are discrete events that occur in continuous time. Specifically, a suseptible individual may be converted to an infected one at a probability which is the sum of the force of infection from all other infected individuals, weighted by a disease disperal kernel. For any individual, infection is a Poisson process, where infection occurs at random times driven by a mean rate, as is mortality, which occurs at fixed rate for infected individuals. At the population level, the dynamics of the disease can be approximated by a set of differential equations:

\(\def\CSI{\mathcal C_{SI}} \def\dCSI{\dot{\mathcal{C}}_{SI}}\)

\[\begin{aligned} \dot S &= -\beta (1 + \CSI) SI - \alpha S \\ \dot I &= \beta (1 + \CSI) SI + \alpha S - \mu I \\ \dCSI &= f(S, I, \CSI, \ldots) \end{aligned}\]

Here \(S\) and \(I\) are the populations of suseptible and infected individuals, \(\beta\) is the overall contact rate between them and \(\alpha\) the background infection rate from exogenous spore pressure, and \(\mu\) is the mortality rate for diseased plants. \(\CSI\) is a spatial covariance term. It is essentially the integral of the distribution of distances between pairs of plants, weighted by the dispersal kernel of the disease.

The evolution of \(\CSI\) can not described by a closed-form equation. It must be simulated at the individual scale and depends on the particular arrangement in space of individuals. However, using the equation-free framework, we can construct a coarse-time stepper that estimates \(\Delta S, \Delta I\), and \(\Delta \CSI\) over short intervals.

Control of the spatial point process

In addition to the internal system dynamics, we also introduce a control variable, \(h\), which represents a harvest of plants:

\[\def\CSI{\mathcal C_{SI}} \def\dCSI{\dot{\mathcal{C}}_{SI}}\]

\[\begin{aligned} \dot S &= -\beta (1 + \CSI) SI - \alpha S - h S \\ \dot I &= \beta (1 + \CSI) SI + \alpha S - \mu I - h I \\ \dCSI &= f(S, I, \CSI, h, \ldots) \end{aligned}\]

Managers may harvest plants in order to reduce rate of disease-induced mortality. Plants provide an ecosystem service continuously (\(\gamma\) per plant per unit time) as long as they are alive, which managers want to maximize. However, the only tool they have at their disposal (in this sylized model) is harvesting the plant population to reduce contact rates, irrespective of plant disease status. Harvesting occurs at a rate \(h\), the portion of plants removed per unit time, and the cost are \(c(S+I)\), where \(c\) is the cost of removing a plant.

The (current-value) Hamiltonian equation of this system is:

\[\begin{aligned} \mathcal H = &\gamma S + \gamma I - chS - chI + \lambda_1 (-\beta (1 + \CSI) SI - \alpha S - h S) + \\ &\lambda_2 (\beta (1 + \CSI) SI + \alpha S - \mu I - h I) + \lambda_3 (f(\cdot)) \end{aligned}\]

The final term, \(\lambda_3 (f(\cdot))\), represents the shadow price of the spatial correlation, or the value of having the individuals arranged in their particular configuration. We can determine the evolution of the shadow prices by differentiating \(\mathcal H\):

\[\begin{aligned} \frac{\partial \lambda_1}{\partial t} = -\frac{\partial \mathcal H}{\partial S} - \delta \lambda_1 &= -\gamma + ch -\lambda_1 (-\beta (1 + \CSI) I - \alpha - h - \delta) - \lambda_2 (\beta (1 + \CSI) I + \alpha) - \lambda_3 \left(\frac{\partial f}{\partial S}\right) \\ \frac{\partial \lambda_2}{\partial t} = -\frac{\partial \mathcal H}{\partial I} - \delta \lambda_2 &= -\gamma + ch - \lambda_1 (-\beta (1 + \CSI) S) -\lambda_2 (\beta (1 + \CSI + \delta) S - \mu - h) - \lambda_3 \left(\frac{\partial f}{\partial I}\right) \\ \frac{\partial \lambda_3}{\partial t} = -\frac{\partial \mathcal H}{\partial \CSI} - \delta \lambda_3 &= -\lambda_1 (-\beta SI) - \lambda_2(\beta SI) - \lambda_3 \left(\frac{\partial f}{\partial \CSI} + \delta \right) \end{aligned}\]

(With \(\delta\) being the discount rate)

The derivatives of \(f\) in these equations can be approximated if our forward integration scheme estimates the second derivative of \(\CSI\), which it will if the forward integrator uses a polynomial scheme.

\[\frac{\partial f}{\partial S} = \frac{\partial f / \partial t}{\partial S / \partial t} \approx \frac{\Delta \dCSI}{\Delta S}\]

For any of the derivatives above, though, we require \(h\), which can be calculated using the maximum principle:

\[\begin{aligned} &\frac{\partial \mathcal H}{\partial h} = -cS - cI + \lambda_1 S + \lambda_2 I + \lambda_3 \left(\frac{\partial f}{\partial h}\right) = 0 \\ &\frac{\partial f}{\partial h} - \frac{S(c - \lambda_1) + I(c - \lambda_2)}{\lambda_3} = 0 \end{aligned}\]

Here we must approximate \(h\) by using a Newton-type method where we calculate \(\partial f / \partial h\) by finite differences, repeated calling the time-stepper. This may be the most numerically costly and challenging component of the scheme, especially when the time-stepper is stochastic.

With the above calculations, we can wrap the Hamiltonian set of ODEs, in a differential equation solver, and additionally in a boundary value solver to determine optimal paths.

I note that this scheme requires that the IBM have an explicit implementation of a “lifting function” for the harvest variable. This could take many potential forms which would correspond to different harvest “strategies”. That is, there are many ways that plants could be harvested from the spatial point distribution for any given rate: randomly, in a fixed or responsive spatial pattern, etc. This optimization can only find the best harvest path for any given “strategy”.

Generalization

This approach should be applicable to similar systems where there is a tractable profit function at the coarse scale \((\pi)\) but where the evolution of the system can only be simulated by a “black box” function \((f)\).

\[\dot S = f(S, h)\] \[\mathcal H = \pi(S, h) + \lambda f(\cdot)\] \[\frac{\partial \lambda}{\partial t} = -\frac{\partial \pi}{\partial S} -\lambda \left(\frac{\partial f}{\partial S} - \delta \right) \approx -\frac{\partial \pi}{\partial S} -\lambda \left(\frac{\Delta \dot S}{\Delta S} - \delta \right)\] \[h^* = h \text{ given } \left\{\frac{\partial \pi}{\partial h} + \lambda \frac{\partial f}{\partial h} = 0\right\}\]

Questions and Next Steps

  • Has something like this been implemented before? I think the multi-scale approach combined with optimal control such as this is novel, but I would not be surprised if a numerical approach had been taking to solving adjoint and maximum equations before.
  • I’m working on implementing a coarse time-stepper for the spatial point process. The thorniest part here is generating a sample spatial point fields from a given \(\CSI\) value. This will probably be the computational bottleneck so I’m working on an efficient approach for it. From there I’ll work on the algorithm from the inside out.


← How to format plots for publication using ggplot2 (with some help from Inkscape) | All posts | Ryan Peek on Creating Shiny Apps →