pdg-control, Training Problem 2 update to group

Intrinsic Stochasticity

\[ x_{t+1} = z_t f(x_t, h_t) \] \[ max_{h_t} \textrm{E} \left( \sum_t \Pi(h_t, x_t) \delta^t \right) \] Measurement error $ m_t = z_m x_t $ Implementation error $ i_t = z_i h_t $

Next: Uncertainty = we can learn and hence decrease the uncertainty. But state-space grows exponentially.

Parametric Uncertainty

Biological state equation \[ x_{t+1} = z_t f(x_t, h_t, a_t) P(a_t, \hat a_t, \sigma_t) \] Bayesian Learning state equations \[ \hat a_{t+1} = \textrm{Posterior}(a | x) = \frac{ L(x|a) \textrm{Prior}(a)}{\int L(x|a) \textrm{Prior}(a) da } \]

Myopic learning just evolves the original SDP problem over the stopping time, integrating over the fixed uncertainty, but updates only to the next time-step. It then implements the optimal next step, observes the system, updates the uncertainty, and then repeats the SDP over this updated uncertainty. This scales as \(n^2\) times the action-space, since the SDP need not consider possible belief states.

Adaptive learning includes the belief states in the state dynamics. Even summarized as a distribution with two summary statistics, this involves a transition matrix of $n^6 $ entries for each possibility in the action-space.

Model Uncertainty

Parameters of the models are fixed, and becomes parameter uncertainty, Updating rule is \[ x_{t+1} = p_t f_1(x_t) + (1-p_t) f_2(x_t) \] \[ p_{t+1} = \frac{p_t f_1(x_t)}{ p_t f_1(x_t) + (1-p_t) f_2(x_t) }\]

State uncertainty

\[f(x,h) P(x,\mu, \sigma) \] Belief MDP or POMDP.

Misc Notes

  • Why is an Allee threshold different than the zero boundary in the other models? Stochastic rescue.

  • Viable vs optimal control (Doyen) (Martinet & Doyen, 2007)

References