Troy noted that he is considering changing the way input data is brought into the model. In the prototype model Jim Norris has been using for model comparisons, the input data are stored and manipulated outside the model in Access databases. Using this type of input format may reduce the amount of code we have to write and helps to meet one of the project objects, namely interfacing with "outside" databases.

A specific problem Troy is currently working on is the shaker algorithm. The old code has only four cohorts per stock--one for each age class. The new model format will have an unlimited number of cohorts. After some group discussion to clarify the problem, it was decided that there would not be a problem if the current algorithm was used to compute shakers for each stock and age class within a timestep and region, and then those shakers were further allocated to the individual cohorts within each age class based on their relative abundances within that region at the start of the timestep.


3. Update on model comparisons (Jim Norris).

Jim briefly reviewed the different migration assumptions, algorithms, and estimation methods used by the three models in question (FRAM, PM, and SSM) and the model comparison approach he proposed at the last meeting. He stated that after the discussion at the last meeting, he decided that the model comparison program should be re-written using the object oriented capabilities of Visual Basic. This would serve two purposes. First, it would provide a prototype within which code design for the new model framework could be tested. Second, it would facilitate a more structured method of comparing the models.

The new model comparison code has a "ModelEvaluator" class that contains several instances of the "Model" class: a BaseModel, a FRAMModel, a PMModel, and a SSMModel. Each instance of the Model class has a "State" class which contains a list of Region, Stock, and Fishery objects. Each Model class object also has several process manager objects (e.g., NatMortManager, MigrationManager) that perform operations on the State during each timestep.

For model comparison analysis the BaseModel's state and process managers are configured to simulate the true state of nature. A daily timestep is used and each process can be simulated using a different algorithm each day, or group of days. The data used to configure a model are stored in Access database files. The user interface allows the user to specify which database files to use and to specify how much variability the different processes will have (e.g., none, low, med, high).

When the BaseModel is initialized, the user interface displays schools from the unmarked and marked cohorts of a stock as tic marks along a linear migration path. The height of each tic mark represents the abundance of that school. Vertical lines along the migration path delineate region and fishery locations. When the BaseModel is run, the school tic marks change position each day along the migration and change height as they suffer natural and fishing mortality. Cumulative catches in each fishery are shown also.

Each instance of the Model class has a DataManager object that keeps a list of true catches and escapements. When the BaseModel run is completed, the user clicks a "Measure Data" control button that directs the DataManager to create a list of measured catches and escapements. The amount of measurement error (none, low, high) is controlled via the GUI.

The ModelEvaluator class has objects that contain methods for analyzing measured data created by the BaseModel. For example, when the user specifies the data aggregation resolution (i.e., number of days per aggregation period) and clicks the "Estimate FRAM" button, the FRAMAnalyst object estimates FRAM parameters and stores them in Access data files. When the "Run FRAM" button is clicked, the FRAMModel is configured, initialized, and run and displayed in the GUI. When the "Run Estimation" button is clicked, the BaseModel and FRAMModel are run simultaneously so the user can see more clearly how the FRAM model simulates the true state of nature.

At present the code for the PM and SSM models are not in the program. Jim felt that he could implement the PM code without too much trouble, but thought implementing the SSM estimation code would be too difficult. Initially, he plans to send "measured data" output to Ken Newman via the internet, have Ken estimate parameters, then use those results to configure the SSMModel to run simultaneously with other models in the model comparisons program. Jim Scott requested that Jim still try to integrate Ken's estimation code into the model comparisons program.

There was a brief discussion of how to simulate estimating FRAM and PM parameters from multiple years of data instead of a single year. Rich Comstock said that the PM model uses the median of the estimated values. Jim Scott and Gary Morishima stated that FRAM uses a weighted sum of CWT recoveries over several years to create a single "base year" of CWT data, where the weights for each year are the relative abundances of each year class.

Jim Norris raised the question: How are the harvest rates used in the PM model estimated? He noted that the PM model does not have a specific estimation procedure for harvest rates. Jim said that for the model comparison program he planned to use either the FRAM harvest rate estimates, or to give PM the "true" harvest rates used by the BaseModel. Another option would be to use the SSM estimates.

Estimating the migration mapping between fisheries for the PM model also was discussed. Jim Norris said he planned to have the MigrationManager in the BaseModel keep a record of when fish moved from one fishery region into another. These data would provide the "true" migration mapping information. Gary Morishima suggested that the PMAnalyst object could estimate migration speeds from the measured BaseModel data, from which a migration mapping could be estimated.


4. Elements of decision theory and relevance to preseason planning (Ken Newman).

Ken reviewed the basic elements of decision theory:
-- action space (e.g., a set of choices for management plan);
-- possible states of nature (e.g., survival rates, migration patterns);
-- loss function (a function of the action taken and the particular state of nature).

Ken showed a simple discrete case example to illustrate two types of decision rules:
-- minimax rule (i.e., choose the action which has the smallest maximum loss);
-- Bayes rule (i.e., choose the action which has the smallest expected loss).

Ken stated that the three most difficult issues in applying decision theory to salmon fishery management are:
-- selecting a loss function (e.g., how to resolve conflicting objectives of escapement levels, catch quotas, and catch allocations);
-- high degree of uncertainty in the states of nature;
-- virtually infinite number of possible actions.

To solve the 2nd issue, for a given fishery management plan (action), one can use the hierarchical state-space model to simulate different state-space model parameters. Given these parameters, simulate the abundance and catches, calculate the loss function for each simulation, and average the losses to get a measure of the expected loss.

 For solving the 3rd issue of an infinite action space, to finding an "optimal" action, Ken suggested that a technique called "simulated annealing. The basic idea of simulated annealing is to start with an input action (i.e., management plan expressed as a set of fishing effort matrices) at time t, call it a(t), and then simulate the plan N times. From these N simulations calculate the expected loss function and let this function be the objective function to be minimized. Then randomly perturb the a(t) slightly to get a(t+1), simulate the new plan N times, and compare the two objective functions. Choose the best action (i.e., plan with smallest objective function) and repeat the process until no better action can be found.

There was considerable discussion about this issue. Gary Morishima wanted to know why we were even considering this type of problem. Jim Norris explained that in order to simulate management actions over many years (a NMFS objective for ESA type analyses), the computer program would require some method of simulating management decisions that currently are handled within the political arena during the preseason planning process (e.g., trying to simultaneously satisfy multiple catch allocation requirements and multiple escapement goals).

Jim Scott pointed out that the current version of FRAM has an algorithm for computing terminal area catches such that escapement goals and catch allocations are satisfied. In essence, this algorithm subtracts the escapement goals from the terminal run sizes entering the terminal area and then allocates the available catch among Treaty and non-Treaty terminal area fisheries. This algorithm operates over several timesteps.

It seemed to be the consensus of the group that there needs to be a distinction between using an algorithm to (1) find solutions to implement a given set of decision rules, and (2) searching for an "optimum" harvest plan.

Ken's handouts on this subject are available at: /newman/dectheory.pdf


5. Estimating migration parameters in more complex spatial frameworks (Ken Newman).

Up to this point Ken's SSM has used only a single line segment to describe a salmon migration path. While this may be adequate for coastal stocks, it is not adequate for stocks entering inside waters, such as Puget Sound or Georgia Strait. Ken described a theoretical probability model to handle this more complex spatial framework. The basic idea is to describe the position of an individual fish as a 2 dimensional random variable [x(t), p(t)], where x(t) is a label for a region the fish is in at time t and p(t) is a particular location within x(t). To implement this type of model it will be necessary to develop a set of "stock rules" describing the possible regions each stock will potentially migrate through. The details of the model can be found in Ken's handouts on this topic, which are available at:

/newman/extensions.pdf


6. Combining multiple effort measures (Ken Newman).

Ken made the following "first cut" recommendations regarding combining multiple effort measures.
-- Given two fisheries with differing temporal resolution: partition the coarser data to the finer resolution, using smoothing techniques;
-- Give two fisheries with differing spatial partitioning, one of which is nested in the other: partition the coarser resolution fishery data to the finer scale, using smoothing techniques;
-- Given two fisheries with differing spatial partitioning, neither of which are nested in the other: assuming one is somewhat coarser than the other, first smooth the coarser fishery data and then partition the smoothed values according to the finer spatial resolution;
-- Given two fisheries with differing temporal resolution and non-nested spatial partitioning: first temporally partition the fishery with coarser temporal data, then spatially partition the fishery with coarser spatial data, in both cases using smoothing techniques.

Further comments on this subject are available at:

/newman/extensions.pdf


7. Applying the SSM model to chinook salmon (Ken Newman).

When applying the SSM to coho salmon, it is assumed that all individuals within a cohort mature during the same year. This assumption will not be valid for chinook salmon. Ken made several suggestions for modifying the SSM to include a maturation process. These suggestions can be viewed at:

/newman/extensions.pdf


8. Data issues (Robert Kope, Jim Scott).

Robert reported that he submitted a letter to the co-chairs of the PSC Technical Committee on Data Sharing (Norma Jean Sands and Susan Bates) requesting that latitude and longitude values be assigned to all recovery site locations used in the CWT database. Norma noted that she had not received the letter yet.

Jim passed out a table showing the different CWT sampling period types used by different agencies in the PSMFC database:

ADFG
-- Commercial Troll and Net: Reported by statistical week (Sunday start). Sample fraction is estimated by open period;
-- Sport: Marine data reported biweekly; Freshwater (?)

CDFO
-- Commercial Troll and Net: Reported by statistical week (Sunday start). Weeks for Jan, Feb, & Dec recoded as "40." 1981 data has a Monday start.
-- Sport: Marine reported by calendar month; Freshwater (?).

WDFW
-- Commercial Troll and Net: Reported by statistical week (Monday start), except troll data prior to 1990 reported by statistical month.
-- Sport: Marine and freshwater reported by statistical month.

ODFW
-- Commercial Troll and Net: Reported by statistical week (Monday start);
-- Sport: Marine and freshwater reported by statistical month.

CDFG
-- Commercial Troll: Reported semi-monthly;
-- Sport: Marine reported semi-monthly; Freshwater (?).

Jim also reported on several other questions:

Q1. How is effort reported in the Canadian database?

In the Canadian database commercial effort and CWT recoveries are generally reported by week (Sunday start), and the first Saturday of the year by definition terminates week 1. All weeks in Jan, Feb, and Mar are recoded to occur in a single period. Depending on source of the data, the week may be reported in the form MMW (e.g., 061 for the first week in June), or consecutively from 1 through the last week in the year.

Q2. How is effort computed by time period for freezer boats in the troll fishery?

Recall that the catch and CWT recoveries from freezer boats are reallocated based upon the catch in the day and ice boats. Unfortunately, CDFO does not currently maintain a database with a similar reallocation of fishing effort. Estimates could be obtained easily from the reported catch and effort of day, ice, and freezer boats.

Q3. Are species specific effort data available for Canadian fisheries?

Jim was unable to locate any species specific effort data for Canadian fisheries.

Q4. How are species-specific effort data computed for Washington fisheries?

Species-specific effort data are computed in two steps. First, fisher or boat numbers are used to identify unique landings within a day. This eliminates the potential confounding effects of multiple landings by a single fisher within a day. Second, only the unique landings that sold fish of the species of interest are added to the effort that the specific effort. For example, if there were 20 unique landings, but only 19 sold coho salmon, the coho-specific effort would be 19.

Jim's major conclusions are:

-- Need to re-compute the effort for freezer boats;
-- Species-specific effort is not very useful and we need to use total effort.


9. Work schedule (Jim Scott).

Jim organized four current tasks as follows:

Task 1. Model Code (Troy lead). Need to implement the harvest code, including shaker and CNR mortalities and different policy types (e.g., selective fisheries, escapement goals, catch allocations). The prototype will duplicate the PSC chinook model code without a GUI. It will include escapement goal management, but not selective fisheries or CNRs. Three more months are need to finish this task. Jim Norris will organize a harvest algorithm work group to help resolve algorithm questions (meeting date set for June 16) with possible participants: Jim Norris, Jim Scott, Gary Morishima, Jennifer Gutman, Jim Packer, Robert Kope.

Task 2. Model Comparisons (Jim Norris lead). Need to get a fully functioning simulator completed and develop an experimental design for comparing models.

Task 3. Data Issues (Jim Scott, Jennifer, Cara, Din, and Carrie will coordinate). Need to get Humptulips data and Voight's Creek data for 86-91. Jim Norris agreed to send a tag code list to Cara. Din will get tag codes for Big Qualicum stock. Jennifer will provide NWIFC mapping of PSMFC database recovery codes to aggregated recovery regions for analysis.

Task 4. Spatial Complexity (Jennifer and Din). Ken's method of estimating migration paths for "inside" stocks will require "stock rules", or migration path definitions, for these stocks. Jennifer will create a set of rules for Voight's Creek; Din will create rules for Big Qualicum.


10. Meeting Schedule and Milestones (Jim Scott).

The following meeting schedule and milestones were established:

July 2 (Thursday):
-- Humptulips results;
-- Model code recommendations from the work group;
-- Model comparison program ready;
-- Experimental design recommendation for model comparisons.

July 30 (Thursday):
-- Model comparison results;
-- Voight's Creek and Big Qualicum results;
-- State space model preliminary decision (will it work better?).

August 27 (Thursday):
-- Look at other coastal stocks;
-- Another review of stock rules;
-- State space model decision refinement.

October 6 (Tuesday):
-- Code prototype completed;
-- Results for many stocks and years!