NOT TO BE CITED WITHOUT PRIOR REFERENCE TO THE AUTHORS
ICES CM 1998/J:11 Session (J): Variation in the Pattern ofFish aggregation: Measurement and analysis at different spatial and temporal scales and implications
Using local and scientific knowledge to predict distribution and structure of herring shoals Steven Macldnson! and Nathaniel Newlands (Fisheries Centre, 22 04 Main Mall, University of British Columbia, Vancouver, Canada, V6T 1Z4,: corresponding author:email:
[email protected],phone: (604) 822-2731,fax: (604) 822-8934)
, Abstract Attempting to predict the distribution and structure of herring shoals is an ominous task given our incomplete understanding of ecological mechanisms. Despite recent attempts to link cross scale behaviour dynamics and distribution studies, there are still large gaps in our basic knowledge of shoal movements. Nonetheless, the knowledge of fishers and fishery managers is not incorporated in to our scientific analysis but such information is rich in observation since knowledge of fish behaviour and distribution is a pre-requisite for their profession. Combining such observations with more conventional scientific studies and theoretical interpretations provides a means by which, we may bridge some gaps in our knowledge. We present a framework in the form of a fuzzy logic expert systemwhose knO\vledge base incorporates both local and scientific knowledge in the form of heuristic rules, The knowledge base is flexible in the ,sense that it can easily be modified to add additional information or change current information. Additional 'operational rules' govern how the system utilises, its knowledge by setting priorities to certain factors under specific situations. The fuzzy system is driven by responses from an external user who provides information on biological and environmental aspects pertaining to the locality and time of year in which they wish to learn about In response, the system predicts structure and distribution patterns of herring shoals. It is context sensitive and 'intelligent', asking the user the minimum questions required to make its predictions. Hypertext screens ". and-graphical- mtert"ace providesuser-rnindl:Y·inriufaJi.oorifpiit 'toileiliei\Vifh-the-capacit)'-to-fiilly--" explain facility of how the system arrived atits conclusion.
Keywords: Heuristics, fuzzy logic, model, prediction, herring, school, distribution
1
Introduction On broad geographic scales, herring populations world-wide display remarkable pattern in their spatial and temporal distribution. The annual migration paths of migratory stocks may cover 1000's of kilometeres between spawning, feeding and overwintering sites, many of which have been known for quite some time (Harden-Jones 1968, McKeown 1984). As ecologists, our challenge is to fathom how such repeatable pattern can be derived from the complex, day to day behaviour of individual herring. It is an onerous task; one that is matched in complexity by the tremendous inherent variations within and between our scales of observation. To develop spatially explicit predictive models needed for management and to allow us to respond to change, we must learn how to interface disparate scales of interest and understand how information is transferred from fine to broad scale and vice versa (Levins 1992). A key element guiding our research on the distribution of herring, focuses on understanding how selection has favoured schooling as an adaptive behaviour. Experimental behavioural studies (review by Pitcher 1993) together with detailed in situ acoustic observations (e.g. Gerlotto et al. 1994 Masse et al. 1996, N0ttestad et al. 1996, Pitcher et al. 1996, Petitgas and Levenez 1996, Maraveliasand Reid 1997, Misund et al. 1997) offer insight on which to base our ecological interpretations. The structure and distribution of herring shoals is currently believed to be governed by the strive to achieve 3 main goals; survival; growth arid reproduction (Ferno et al. 1998). Previously, most attentibnhas been directed to small scale (0.1 to 10 m, seconds to minutes), school organisation and dynamics; or, large scale (100' s km, weeks to months), stock structure and migration studies. Information on the mesoscale (0.1 to 100km, day to weeks) distribution pattern of schools and school clusters is particularly lacking. . Since many herring fisheries are typically conducted at spatial scales of one to tens of kilometres and occur for periods of days to weeks, both fishers and fishery managers alike operate within themesoscale realm. By virtue of their profession it is prerequisite that they have knowledge regarding the distribution and behaviour of herring. Such a rich information source should not be unutilised. Typically however, local ecological knowledge has been considered by moretniditional analytical fisheries science as 'anecdotal', and for the most part been absent from fish stock assessment or during development of management' plans. In contrast, social science emphasises the importance of local knowledge in a contextual and historical perspective, but again for the most part, has been unsuccessful in directly incorporating this informatiori into fisheries management (see Neis et at 1996 for exception). The relt!ctance and inability to,lltilise J1911-scientificknoWledgeasdata is asignifiqmt affliction hindering the progress of fisheries science and management (Mackinson and N0ttestad, in press). There is an urgent need to maximise the usefulness ofinformation by incorporating alternative sources of data and finding new ways of using traditional ones. We present a formal framework for combining local ecological knowledge and scientific knowledge in the form of heuristic rules and demonstrate how it can be used to predict the structure and distribution of adult herring shoals for different phases during their annual life cycle.
Heuristic Data Sources (developing the knowledge base) This heuristic model is developed within the framework of an expert system (Figure 1) and incorporates two fundamental data sources of information on fish distribution and behaviour; (i) 'practical' data: local knowledge of interviewed fishers, fishery managers and First Nations people; (ii) 'hard' data: scientific information from interviewed fisheries scientists, field work studies and published literature sources.
2
Briet1y, expert systems are a branch of artificial intelligence; theories .and methods for automating intelligent behaviour. They are computer programs that provide assistance in solving complex problems normally handled by experts. They use rules to· store. knowledge. When the system is asked to solve a problem, it uses this knowledge to infer solutions. The software used to develop the system presented here is Exsys®' Professional by Multilogic. Figure 1. near here
Interviews A total of 30 formal interviews were conducted, half with fishery scientists (8) and fishery managers (7) who generally provided technical or hard scientific knowledge, and half with fishers (9) and First Nations people (6, all previous or currently herring fishers) who provided professional practical local knowledge. With the exception of one gillnetter who specifically undertook herring surveys, most fishers interviewed were seine fishers, with a collective experience (CE) of approximately 270 fishing years. Seiners were specifically chosen as candidates in contrast to gillnetters or roe-on-kelpfishers, since seine-fishing typically requires specific knowledge of fish distribution and movements when searching broad areas. The First nations CE amounted to approximately 290 fishing years. Selection of interviewees was deliberately non-random. An attempt was made' to interview those fishers who had the most experience fishing herring during different seasons, at different locations and who were held with respect by other fishers in the community. For this reason, progressive selection of interviewees was conducted by word of mouth, one candidate suggesting another to talk to. This method proved to be very successful. Fishery scientists and managers were selected based ontheir experience with herring. The current regional herring co-ordinator and 3 long time fishery managers (CE approx. 160 years) offered a more technical 'field based' perspective that complimented observations by fishers. Three herring scientists from the Pacific biological sta.tion, Department of Fisheries and Oceans (CE approx .. 75 years), and a further 5 from Norway (Institute of Marine Research and University of Bergen; CE approx. 80 years) provided hard scientific observations from field and experimental studies. Typical interView duration was 2 hours but some ranged from 1- 4. With two exceptions, all interviews were conducted on a separate individual basis at the preferred location of the candidate. For help with interpretation, it was necessary on one occasion to interview two fishers together. In another . instance, a meeting was held withtl1e. first Nations Sliamm()nband.elders, that included men and. women who had traditionally been involved in herring fishing prior to the demise of their localised hei:ring stock. . . Two basic questions were asked; what? and why? Firstly, interviews were asked to recount what theY'had observed regarding distribution and behaviour of herring and secondly to offer possible explanations to account for their obserVations. All candidates were asked the same type of questions although specific interViews were 'free range' or 'adaptive' in the sense that the format and directness in which the questions were presented depended upon the context of discussion. Allowing discussion to cohtinue openly in this manner prOVided insight in to many aspects which would have been overlooked by a simple questionnaire offering only a fixed set of responses. On almost all occasions, completely new knowledge was acquired. Using the same technique of Johannes (1978), reliability of the subjects' answers was tested by asldn,g two tyPes ofquestions at a~_onYenil:mJ tilllecluDngcliscussiQn.. Ihej'irst""ere question.s tow.hjch._. the answer was already well-known. Responses to these questions were almost always the correct answer or that they didn't know. The second type of question sounded plausible, but which were ones that the fishers were unlikely to be able to answer. In virtually every instance, the response to this type of question was "I don't know", although on two occasions plausible explanations were offered.
3
Interviews were recorded by hand written notes and subsequently typed and resent to the candidate within 48 hrs for verification of accuracy, ;corrections and additions. Prior to. interviewing, candidates signed an ethical review form affirming that the information received would remain their property and that reference would be given directly to them when cited. At a later stage, details ofal! candidates together with a full account of their interviews was recorded in a database. In addition to interviews, 18 questionnaires were posted bye-mail to researchers involved with herring populations world-wide. Despite a second reminder, there was only one response that contributed to the knowledge base. There was clear demarcation in the type of responses from different interviewees. Typically fishers were particularly strong .on observation providing detailed accounts of school structure, distribution and behaviour including; school sizes, shape, density, depth distribution, association with specific features, ease of capture and specific behaviour patterns relating to season, tide, weather, fishiI).g vessels, time of day, feeding, occurrence of predators. However, when asked why;. they were generally reluctant or found it difficult to offer an interpretation. It was difficult to elicit a rank order of factors they· considered important in determining the observed shoal structure, distribution and behaviours. In contrast, fishery scientists were more familiar and at e.ase with offering. interpretations for their .observations or experimental findings and for the most part, were able to assign an order of relative importance to the factors contributing to shoal structure, distribution and behaviour patterns. Responses of fishery managers were more akin to those of fishers, being grounded firmly in field observations. Due to the nature of their job, most were uncomfortable with ascribing behaviours to any particular factor. They tended to err on the side of caution and uncertainty, usually offering provisos and comments of exception to any of their observations. They were however, more willing than fishers to offer potential interpretations, and it was apparent that these were frequently guided by scientific understanding from colleague fishery scientists. Remarkably there were no instances. in which knowledge accumulated from any single source diverged. Information either complimented previous knowledge or added additional understanding. Moreover, the majority of comments contained information pertaining to mesoscale spatialanc\ temporal distribution pattern of schools. Field work and literature sources Information from four research cruises is compiled within the knowledge base.. Two cruises were cond)lcted in Pacific coastal waters of British Columbia during pre-spawning and spawning period; one in the Strait of Georgia (1997) and another in the Central Coast (1998). A specific attempt was made in .the 1998 survey to. valicjate observations from fishers. They were used to form hypotheses on which the survey was analysed (Mackinsor\!n prepfTwo further cruises werecondilci:edm-theNorwegian sea during May 1997 and 1998, during a period of migration and ocean feeding for Norwegian spring-spawning herring (see Mackinson et al. this conference for further details). Literature sources concerning· field and experimental studies on behaviour and .distibution of herring and other schooling pelagic fish was reviewed. Relevant information was recorded directly in a database cross-referenced to its original source. The 'hard' data obtained from field work and 1iteratur~ sources was used for defming ranges for descriptors of structure and distribution .. Through the process of defuzzification these .values are combined with information from heuristic rules to provide numeric output from the system. This is described in detail in a later section.
Structuring tbe heuristic system (k.nowledge Imgineering) .. Structuring of the heuristic system involves the process of building a multi-layer decision tree, that relates information from different spatial and temporal scales. After considering alternative strategies for dealing with information, it was deemed most appropriate that all knowledge, would be 4
contribute equally in building the knowledge base. Therefore, an assumption is equality in the degree of belief in a piece of information from either fishers, fishery managers, fisheryscientists, First Nations people or from literature. This assumption helped to maximise all of the data sources (Mackinson and N0ttestad, in press). Combining knowledge using heuristics Knowledge from interviews and literature was sorted in detail to identify information relating descriptions of behaviour and distribution pattern (descriptors) to those factors having influence upon them (attributes) A complete classified list of the extracted descriptors and attributes is provided in Figure 2. Figure 2. near here Using a series of heuristic rules capturing knowledge, the relationships between attributes and descriptors were defmed in .the expert system. Many of these rules came directly from source whilst others were defined on best inference. A rule has the form IF this THEN that and may contain .several conditions in the IF part linked by AND, OR, NOT, and one or more .elements in the THEN part linked by" an AND. Each item contained in the THEN part is then assigned an associated confidence.value according to the relative contribution of information relating to that particular rule. For the most part, attributes are used in the IF part and descriptors are used in the THEN part of rules since the goal is to conclude upon how behaviour and distribution is determined by the influence of various combined attributes. A example rule is:
IF type of predator is an aquatic predator (fish or mammal) AND abundance of aquatic predators is high THEN' packing density of school is high (item conf = x) AND internal dynamics schooling (item conf= y) Within the expert system, attributes and descriptors are classified according to their properties (strings or variables), in the way they are used in rules, and whether they are required as input or output to the expert system. There are two basic types with subdivisions: (i) Qualifiers (lists of options/ choices to chose from) may be designated as single choice, multi-choice or as fuzzy sets associated with a variable, (ii) Variables may be defined as numeric, text or strings but are most commonly applied as numeric variablesusedin association for outputting descriptors. In the example-rille above,type of predator operates as. a multi-choice input with two options,. predator type is either 'aquatic predator' or 'avian predator'. Abundance of aquatic predators is designated as fuzzy sets associated with a fuzzy numeric variable, 'predator abundance scale', Figure 3 . . The use of fuzzy sets as qualifiers allows us to build a system that. directly represents linguistic expressions containing useful and relevant knowledge. It is unreasonable, impractical andalrnost impossible to relate information in a purely quantitative way without making gross assumptions. More importantly, implementation of fuzzy qualifiers allows us to define relationships between attributes and descriptors on a continuous scale; This aspect provides particular benefit in managing the whole system since it slashes the number of rules required to describe relations. Furthermore, fuzzy rules provide a . direct connection for combining quantitative. and qualitative knowledge and expressing uncertainty. They are the key to achieving quantitative output from qualitative understanding as will be shown later. Interconnected associations between fuzzy qualifiers can be conveniently expressed using a fuzzy associative map (F AM) (Figure 4) Figure 3. near here Figure 4. near here
5
Imposing hierarchy and uncertainty Several techniques are implemented to represent hierarchy in the behavioural decision chain and temporal changes in the importance of factors influencing distribution and structure. We assert that specific behaviours are dependent on seasonal and diurnal changes in motivational state. There are four ways in which the importance of various attributes are altered. 1. Adjusting weights through a series or rules that define seasonal priorities to motivational states of hunger, risk avoidance, reprodliction and energy saving. These are manifest by assigning a weight factor (low, medium, high) that adjusts the influence of particular attributes on the overall outcome. For example, when risk avoidance is a high priority, a high weight is assigned to increas.e the abundance of predators. This is a pseudo-weighting, that is based on the assumption that increasing predation risk can be represented by increasing the abundance of predators. Ideally, weight factors should be directly applied to confidence factors associated with rules, but it is not possible to implement such a method using the current software. . 2. By assigning confidence factors to each rule we influence their importance when fired and propagated through the system. This method imposes hierarchy in the importance of different attributes on descriptors whilst simultaneously representing. the uncertainty. or degree of belief in - the rule. 3. Using operational logic including rules and commands that defme how the system fires under specific circumstances .. In this instance, particular rules may be ignored whilst others are followed. Variables may be assigned values directly as a consequence of operational rules (in particular, when they are deemed oflow importance) 4. Providing the user the opportunity to assign low importance or ignore a particular element. By offering placebos (choices that do not lead to ,any conclusions), such as 'not sure' or 'low importance' the user can exclude or reduce the influence of attributes. However, if the user answers 'not sure' where possible they are gnided by a series of hypertext screens and nues to make a conclusion or decide between ignoring the attribute or letting the system assign a default value.
Predicting distribution and structure When queried, the expert system guides the user through a series of questions, each step comparing the input information with theru\t;s in ihekn.owledge base. OPerationallogicensures.tI1aL each run is unique. Based on previous answers the system prompts the user, asking only those questions pertinent to draw conclusions relevant to· the initial conditions specified by the user (Figure 1). Ru1es using qualifiers defined' as fuzzy sets· are followed in parallel and to' a partial extent since all rules may be true to a certain degree. The operation can be viewed as a parallel mu1ti-layer decision" tree and this aspect separates a fuziysystem from a more simplistic non-fuzzy decision-tree system in which branches of the tree are triggered independently. Propagation of confidence The confidence 'associatedwilh specific qualifiers is set in one of three ways; taken as being 1 when setby the user; determined'by a fuzzy membership function for fuzzy qualifiers; if it is assigned a value in the THEN part, the confidence is propagated from the IF part and combined: with any confidence in the THEN part. Confidence in the IF part The IF part may be made up of one or more conditions linked by AND or OR. Depending on whether the IF part is made up of only one condition or a single qualifier value the confidence is simply the confidence associated with that condition or qualifier. If there are more than one qualifier values .set
6
using an OR such as; IF food abundance almost non OR sparse, we combine the confidence of 'food abundance almost non' and 'food abundance sparse', using the default formula: Combined conf. = combined conf. +(Value conf. *(l-combined conf.))
Eq.l
The formula is known as the MYCIN method (after the first system that used this method) and has the advantages that the confidence never exceeds I, each positive confidence increases the final confidence, and it not sequence specific. The same method is used for combining the confidence of blocks of OR statements. in the IF part of a rule. The overall confidence in the IF statement is calculated as the product of the parts. This simply means multiplying the combined confidence of conditions linked with an OR with those conditions linked with AND. By setting a threshold, rules are only considered to be true when the overall IF confidence is above 0.01. Confidence in the THEN part .. In the fuzzy system, when a rule fires, the IF part will be TRUE, meaning all the parts in the IF statement have a confidence >0. The final THEN confidence is achieved by multiplying the overall IF confidence with the confidence assigned in the THEN part and combining it with the current confidence by a series of formulas (again, the MYCIN method); Current conf. = (IF conf.* assigned THEN conf.) +(Curr. Conf.*(l-( IF conf.* assigned THEN conf.)) Eq.2 The formula has the characteristics that any positive value increases the finalconfi.;lence, and when a THEN part has a low confidence, a high value significantly increases it, but when a THEN part has a fairly high confidence, additional high values only increase it slightly. There is one important shortcoming in the use of the above formulas for propagating confidence. When the IF part has an overall confidence of 1, which occurs when a user directly chooses a qualifier value rather than it being calculated during inferencing,and this is combined with a THEN item whose assigned confidence is also 1, the final IF confidence is pushed immediately to 1. Thus, the confidence combined by firing of other rules has no apparent effect. The simple pragmatic solution used here is to ensure that all· items in the THEN part are assigned confidence values
