Carl Boettiger

Have finally got­ten my active adap­tive man­age­ment solu­tions work­ing. The active adap­tive man­age­ment solu­tion learns very quickly and rather monot­o­n­i­cally which model is cor­rect, even when mod­els dif­fer by small amounts and the ini­tial belief is heav­ily skewed to the wrong model. I’ve imple­mented exam­ples with Myers model, but in this para­me­ter space the allee …

A cen­tral ques­tion of train­ing prob­lem 1b addressed at the work­ing group was how to com­pare the dif­fer­ent penal­ties. Fol­low­ing up on those dis­cus­sions, here are a few notes on how far I’ve got­ten in the analy­sis regard­ing the best way to make these com­par­isons. simulate_npv_curves Nice job of show­ing the per­for­mance of the vari­ance …

Spent way more of today than I wanted to ham­mer­ing out the active adap­tive man­age­ment imple­men­ta­tion for a triv­ial model-choice prob­lem using the dis­cretized belief-SDP approach. We have alter­nate mod­els \(f_1\) and \(f_2\) of the state equa­tions (pop­u­la­tion growth dynam­ics) \[ x_{t+1} = z_t f(x_t) \] and intro­duce a con­tin­u­ously val­ued belief prob­a­bil­ity \(p\) that …

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 \[ …

Triv­ial exam­ple of a non-optimal but more robust pol­icy by edit­ing the opti­mal strat­egy directly. Note in par­tic­u­lar that this out­per­forms the opti­mal strat­egy eco­nom­i­cally as well as eco­log­i­cally (fewer crashed pop­u­la­tions). Run­ning the policy-costs model with Bev­er­ton Holt instead of the Myers func­tion seems to give a clearer pic­ture of a sys­tem respond­ing to …