Recently it has been found that the earth's oceans are warming at a pace that is 40% faster than predicted by a United Nations panel a few years ago. As a result, 2018 has become the warmest year on record for the earth's oceans. That is because the oceans have acted as a buffer by absorbing 93% of the heat produced by the greenhouse gases [1]. The impact of the oceanic warming has already been felt in terms of the periodic warming of the Pacific Ocean as an effect of the ENSO process. The ENSO process is a cycle of warming and subsequent cooling of the Pacific Ocean that can last over a period of years. This cycle was first documented by Peruvian fishermen in the early 1600's. So, it has been part of the environmental challenges that have been presented to economic agents throughout the world since then. It has even been suggested that the cycle has increased in frequency over the years, perhaps in response to the overall issues related to global warming. [2] [3] In this paper Cultural Algorithms are used to develop a multi-objective agent-based model of artisanal (traditional offshore) fishing behavior in coastal Peru, Cerro Azul. The data used to produce this model comes from the observation of fishing behavior over a four-year period, 1982-1986. During this period over 6000 individual fishing trips were documented. This observation period overlapped with one of the largest ENSO activities ever recorded. As a result, it was possible to observe the changes in fishing behavior that were the result of this process. While the data is several decades old, the ENSO process was first observed in Peru in 1502. Thus, this data can be considered to reflect the adaptations that have been made to the process in the ensuing centuries by the subsequent generations to maintain and sustain their well-being.
Evolutionary computation is a subfield of
Artificial Intelligence which is based on
Darwinian principles of evolution. Evolutionary
computation is often applied to the solution of
complex computational problems especially
global optimization problems. Several
Evolutionary Computation systems have been
proposed, one of them is the Cultural Algorithms
[
Recently it has been found that the earth’s
oceans are warming at a pace that is 40% faster
than predicted by a United Nations panel a few
years ago. As a result, 2018 was the warmest year
on record for the earth’s oceans. That is because
the oceans have acted as a buffer by absorbing
93% of the heat produced by the greenhouse gases
[
In Figure 1, sea birds eat anchovetas and deposit nitrogen-rich guano on sea cliffs and offshore islands. Next, humans retrieved guano and use it to fertilize inland maize fields. Much of the maize was converted into chicha in Cerro Azul’s breweries. The Cerro Azul’s nobles lavished chicha on the fishermen who harvest anchovetas for them. Finally, Anchovetas harvested at Cerro Azul were exported to inland communities and paid as tribute to hereditary leaders.
With the onset of an ENSO the anchovies
population shifts to the south towards cooler water
and fish from warmer waters begin to move into
the area from the north. Fisherman must therefore
adjust their fishing strategies to compensate for
this adjustment in the food chain to preserve their
own well-being. So, ENSO has been part of the
environmental challenges that have been
presented to economic agents throughout the
world since then. It has even been suggested that
the cycle has increased in frequency over the years
in response to the overall issues related to global
warming. [
In this paper Cultural Algorithms are used to develop a multi-objective agent-based model of artisanal (traditional offshore) fishing behavior in coastal Peru, Cerro Azul. The data used to produce this model comes from the observation of fishing behavior over a four-year period, 19821986. During this period over 6000 individual fishing trips were documented. This observation period overlapped with one of the largest ENSO activities ever recorded. As a result, it was possible to observe the changes in fishing behavior that were the result of this process. While the data is several decades old, the ENSO process was first observed in Peru in 1502. Thus, this data can be considered to reflect the adaptations that have been made to the process in the ensuing centuries.
The paper is organized as follows: In section 2, a basic overview of the Fishing Trip data set is described. Section 3 briefly describes the contents of the database used to produce the multiobjective model. Section 4 provides an overview of Cultural Algorithms and the particular version, CAPSO, used here to assess the resultant Pareto functions produced by the model. The Artisanal fishing model is presented in section 5. Section 6 provides the resultant Pareto fronts. Section 7 presents the conclusions.
The data used for the analysis here is from the 1980s while the historic site is more than 500 years prior. Drs. Joyce Marcus and Maria Rostworowski led a team of archaeologists from the University of Michigan from 1982 through 1986 to excavate five seasons of research at ancient nearby site of Cerro Azul. Due to arid weather, architecture, fishing nets, the fish middens from 1100 to 1470 A.D were all well preserved at the site. Dr. Marcus explored early issues of “community self-sufficiency” and “community specialization” during Incan times with respect to the site.
The Kingdom of Huarco contained two
localities [
While modern fishermen use equipment that allows them to catch a wider variety of species, the catches can be destined for local consumption or exported commercially to larger cities, such as Lima. As a result, the movements of certain catches that are targeted for commercial sale are more likely to be tracked than others, and fishermen may want to take more risks or more effort to find them. These factors will be key to the model developed earlier.
The first documented instance of an ENSO was in coastal Peru in the early 1500’s. Data from the ancient site of Cerro Azul tells us that local fishing has been a major part of the economy at least 500 years. Thus, the data collected from current artisanal fishermen can be viewed to reflect a long term adaptation to the periodic warming and cooling of local waters in order to continue their reliance on fishing as part of their major local economy. The data subsequently collected by Dr. Marcus was designed to reflect on the issues of economic sustainability and the well being of individual fishermen in the wake of the changes brought about by ENSO/.
In the last three years of their project, Dr.
Marcus began recording the catch of every boat
that returned to the Capitanian del Puerto with the
cooperation of the local government. In addition,
further data on fishing was collected from Peru’s
Instituto del Mar, [
The fishing activity around Cerro Azul is a complex system that has many different parts that interact with each other. We can view the different levels as Macro, Meso, and Micro in terms of their temporal scale. The three basic phases of ENSO constitute the Macro scale. The Meso-scale is represented by the monthly statistics. The micro level corresponds to the days of the week for a given week. These form the basic structure of the Cerro Azul database constructed here.
The fishing activity around Cerro Azul is a
complex system that has many different parts that
interact with each other. One can view the
different levels as Macro, Meso and Micro in
terms of their temporal scale. The three basic
phases of ENSO constitute the Macro scale. The
Meso-scale is represented by the monthly
statistics. The micro level corresponds to the days
of the week for a given week. These form the
basic structure of the Cerro Azul database
constructed here. The ability to investigate the
performance of a complex system as difference
scales of granularity or detail has been suggested
as an important avenue with which to understand
ancient societies as complex systems [
1. Macro-scale provides analytics that summarize behavior over all three phases of the ENSO cycle: Residual El Nino, La Nina, and Back to Normal. 2. Meso-scale corresponds to monthly patterns of behavior. 3. Micro-scale provides statistics about fishing behavior on a daily basis. This provides information about communication of fishing strategies among fishermen in order to maintain an overall level of performance.
These results were then used to constrain the multi-objective model and its behavior. The results reflected the importance of both catch quality (payout) on the one hand and the investment of resources in terms of number of trips and over all distance travelled on the other (sustainability).
One of the key themes that is the learning curve where the catch counts can increase over time during the week relative to certain targeted species. There are indications that there is a priority in terms of what catch to pursue first. Another interesting pattern can be seen with fall back catches which mimic trends in targeted catches. As the number of desirable catches starts to dwindle in an area, the deficit can be made up by Fall Back categories of catches. These catches include sharks, chancho marino (dolphins), among others. The goals of catch quality and trip investment will be key to the multi-objective model that we develop. To prepare for the computational demands of a multi-objective approach an extension of the Cultural Algorithm, will be described in the following section.
Multi-Objective Cultural Algorithms will be
used to validate the agent-based model of artisanal
fishing. If the goals in the model are conflicting,
then one should expect that an optimal decision
represents a tradeoff between them. This will
result in a hyperbolic model for a Pareto front.
Computation of the hyperbolic curves from a set
of examples - is an NP-Hard problem. This
hyperbolic model is then compared to a best fit
linear model to determine which best describes
the simulation results. These results can then be
statistically compared between the curves
generated at each scale. CAPSO is a Cultural
Algorithm Particle Swarm Optimizer. It uses
collected domain knowledge to implement a
parallel recursive search of the problem space
using multiple swarms of agents based upon the
Particle Swarm methodology [
The Cultural Algorithm has three major components: the population space; the belief space; and the communication protocol that define how knowledge is exchanged between the first two components. The population space is defined as a networked set of agents that can provide solutions to an optimization problem.
These individuals are connected by a social fabric over which information can be passed. The belief space can be defined as the collection of experiential and domain knowledge, which can be influenced by individuals within the population space according to their varying degrees of success. The belief space also has the ability to influence following generations of individuals within the population space.
The following is the pseudocode of a generic Cultural Algorithm: 1. The algorithm begins by initializing the
Population and Belief Space. 2. Individuals in the Population Space are first evaluated and ranked through a fitness function. 3. An acceptance function, Accept(), is used to determine which individuals within Population Space will be allowed to update the Belief Space. 4. Experiences of those accepted individuals are then recorded in the Belief Space through the function Update (). 5. The resultant updated knowledge sources then compete and cooperate to produce knowledge driven changes to agent problem solving behavior. 6. Steps 2 through 5 are the evolution loop which repeated until the termination condition is satisfied.
The two feedback paths of information, one through the Accept () and Influence () functions, and the other through individual knowledge and the Evaluate function create a system of dual inheritance of both the population and the belief spaces. The Cultural Algorithm repeats this process for each generation until the pre-specified termination condition is met. In this way, the population component and the Belief Space interact with, and support each other, in a similar mode to the evolution of human culture. A visualization of this process can be found in the following diagram:
The CAPSO system [
The Main function recursively calls SearchInSpace to generate a new swarm thread. A swarm population is associated with that thread via a call to PopSpaceAlg. PopSpaceAlg is in charge of updating the swarm associated with the thread. Each new swarm is given a number of generations to add a new point to the Pareto front, maxGensWoImprov. If it has not by then, it is removed and control is returned to its parent. If it is productive over a maxRepeatsBeforeDivide, it is divided into a number of new subspaces, newSubspace.
In PopSpaceAlg agents are awarded points for
the number of agents currently in the Situational
Knowledge that it dominates in one or more
dimensions. The sum of those points for an agent
is its objective function value. The VegaMethod
(Vector Genetic Algorithms) is called then called
to select the elite points from the swarm.
CASteps is then called and accepts a certain
number of points, the elite, into the Belief Space
in order to update its content. It then applies the
knowledge sources to selectively modify the
remaining threads based upon their relative
performance as expressed in Relative Roulette
Wheel. The process continues recursively until
only one thread remains and is unable to generate
new points in a certain number of generations. In
that case the system can be restarted with a new
random swarm but still using the acquired
knowledge from the currently completed run that
resides in the Belief Space [
A traditional single objective problem is the result of a combination of contributing terms. G1 = P1, P2, … PN where N is the number of contributing factors or sub-goals that are correlated with each other. In a multi-objective problem, the goals can be conflicting and need to be addressed separately. Neither can be completely achieved without some sacrifice with regards to the other. Based upon the prior statistical analysis two basic goals of the Artisanal fishermen were identified:
Goal 1 reflects the need for profitability with regards to the artisanal fishing activity for a given household. If given the opportunity to choose between a catch that can fetch a higher local market value than another, this goal would be in favor of targeting the higher market value catch. To the extent that this can be done over a succession of trips for a family, the presumed social unit here, the fishing agent can even reap a profit over time.
Goal 2 relates broadly to the issue of sustainability and agent well-being. That is, the agent needs to invest sufficient resources into a trip in order to bring back something in order to sustain the family unit and perpetuate the fishing activity. It reflects the general goal of just being able to get out and fish on a given day.
The two formulas of agent-goal achievement are explained next. Goal 1: High Profitability (HP), or Payout for a trip. This is calculated by multiplying the total Fish Count (number of actual fish caught) by the fish desirability. Some fish are more desirable and will sell for much more than others. We use the following formula: Payout = Fish Count * Desirability, where Desirability equal 3 for High, 2 for Medium, and 1 for Low. Goal 2 main objective is sustainability, or Maximum Required Effort expended (MRE). Goal 2 is calculated as: Effort = ((RTD/MPG) * RE) / Fish-Weight, where RTD = Round Trip Distance in KMs from Cerro Azul, MPG = 5 (8 KPG) Relative Effort = 1 for Cerro Azul and 3 = for North/South, FishWeight in Kilograms. The goal of sustainability was expressed in terms of an effort function. The contributors to effort were as follows: 1. Round trip distance (in KMs) between the port of origin, assumed to be Cerro Azul, and the site where the catch was made. 2. Fish weight in (KGs). That will influence the effort taken to transport it back to port. The greater the catch weight the more transport effort needed to do so. 3. Miles per Gallon. While the boats used are all much smaller than commercial vessels, they do vary in size and capacity and therefore require engines with different power requirements. For this model a MPG value that reflects a middle ground in terms of engine power was selected, one that would be a reasonable approximation of a range of motors that the agents might possess. 4. Relative Effort is a multiplier that adds some additional resistance to the journey. If the trip is for Cerro Azul and vicinity, then the round trip distance is not adjusted. If the trip involves travel up or down the coast and away from Cerro Azul a simple multiplier was incorporated to reflect the additional effort that would need to be made in those situations with regards to weather, currents, etc. 5. Fish weight: While profitability was expressed in terms of catch count, effort needs to be expressed in terms of overall weight. For a given trip the weight of the catch was used to predict effort while the count predicted profitability.
The Effort performance function is then simply the distance divided by the miles per gallon times a relative effort booster to reflect other hidden costs, divided by catch weight in KG.
These two potentially conflicting goals are now the basis for the construction of representative trip models or tours. A tour is a sequence of trips that are produced by the concatenation of individual trips that follow a particular set of goal priorities for an agent. That is, what would a series of tours look like if an individual agent had the same goal priority throughout the phase. The strategy used to generate a path through a Trip Graph is determined by a tuple, (HD%, MRE%) that represents the likelihood of preferring the high payoff goal #1 or the minimum required effort goal #2 on a given day for an agent.
For example, (75, 25) means that the likelihood that a profit maximizing trip is selected by an agent is 75/100. If there is more than one trip that has the same HP level, a random number generator then picks the goal for that day which is then used to select the trip. Figure 5 demonstrates how a tour that supports a (HP 0%, MRE 100%) strategy was generated. Figure 6 gives the detail for each of the trips selected to comprise the tour. catch. All of the trips in our database represent successful trips in that regard.
What the curve in Figure 7 means is that many
targeted catches can be found within a short
distance from Cerro Azul during the time of year,
March through June, in Phase 1. This is the
conclusion of the El Nino which is moderated by
the fact that it is the tail end of summer and
beginning of fall. Warm water fish are enticed to
remain in the area even though the warming phase
of El Nino has diminished. It suggests that a
productive sequence of trips in terms of Payout
will be more dependent on timing than on
location. Once fishermen are required to put more
resources into the tour in this Phase, the Payout
drops exponentially [
The agent-based model of artisanal fishing was used to generate tours through the trip graph over a given Phase and a corresponding set of days of the week. A Pareto Curve of nondominated points is produced for each Phase, and each Days of the week scenario. (3 X 4) These curves are now compared graphically in order to identify the decision-making adaptations made by agents to the changing local climate produce by ENSO.
Figure 8 compares the three phases relative to their tours over all days of the week. Phase I: (March through June 1984) lasts only one-third of a year as opposed to the other two. If the same pattern is played out in the missing two thirds of the phase, the maximum total would reach at least that of La Niña. (120,000). Payout declines exponentially with increased effort which suggests that the fishermen did not have to venture far from Cerro Azul in order to achieve the expected payout in El Niño. Sustainable fishing activity could then easily be performed nearby Cerro Azul.
In order to achieve similar total payout (120,000) in Phase II, the fisherman need to take more trips and invest more resources in order to produce a successful trip. There is less of an exponential drop with distance from Cerro Azul in Phase III, so this suggests that they have more experience in fishing in the Back to Normal phase and are able to make better predictions about catch behavior and location.
Figure 9 represents the tours produced
without the consideration of Sundays. Here, the
results are similar to that of the Full Week since
few trips went out on Sundays. In addition, often
there was not an official at the docks to record
trips that went out and came back on Sundays. In
general, Back to Normal again had less of an
exponential drop which suggests that it was an
easier curve to plan for as an agent since the
environment is now more predictable [
Figure 10 represents the fishing behavior exhibited by those fishing early in the week. In Phase I fisherman stayed close to Cerro Azul and took in about half of the Full weeks’ amount. Since it is the beginning of the week the focus is on nearby areas where knowledge of fish locations can be learned. In Phase II the environment is less predictable and fishermen need to travel further at the beginning of the week, than for the other two phases. Likewise, in the Back to Normal Phase III fishing patterns are more predictable as a result of past experience and the slope is very steep since they are able to make good predictions about catch location.
Figure 11 gives the curves for those trips conducted later in the week. What is interesting here is that there is a steep drop in catch quality with increased effort as opposed to the Monday through Wednesday period. This suggests that in all three Phases there was a learning curve such that those trips later in the week benefited from the information collected by trips performed earlier in the week through communication among fishermen. This suggests that there was a informal communication network over which previous trip experiences were distributed. That information, made it easier to plan for later in the week and provided a safety net for decisionmakers.
In addition, a profitability target optimum of 120,000 units emerged. Agents tried to attain that in each phase. However, this was maintained at a cost. The maximum effort needed to produce a successful trip increased from the El Niño to La Niña for example. In order to sustain their fishing endeavor, more resources were invested. Different fishing strategies based upon both phase and days of the week emerged. Clear difference in Monday-Wednesday and Thurs-Sat. tradeoffs were observed. The La Niña Pareto Front had a more gradual decline in profitability with increased effort. This suggests that more care had to be taken in the planning process in order to sustain fishing activities in this transitional phase.
Traditional fishermen in Peru have been adjusting their strategies to changes in the food chain brought about by ENSO activities over the centuries. The goal of this paper is to use AI to understand the nature of strategies used to sustain individual and social well being in the wake of ENSO. In particular, the focus is on what information is being propagated through their social network to support such wellbeing socially and individually.
Our results suggest that indeed the collective economic response of the fishermen demonstrates an ability to adjust to the unpredictabilities of climate change, but at a cost. It is clear that the fishermen have gained the collective knowledge over the years to produce a coordinated response that can be observed at a higher macro level (Pareto Front). Of course, this knowledge can be used to coordinate activities only if it is communicated socially within the society. Although our data does not provide any explicit information about such communication, there is strong indirect evidence that the adjustments in strategy are brought about by the increased exchange of experiences among the fishermen.
The Pareto distributions suggest dominant and successor waves of strategies that may be associated with the length of time over which the simulation window is conducted. These waves represent how subsequent trips were influenced by the knowledge brought back by agents from earlier trips on the days right before that trip. For example, Dr. Marcus recalls one brother talking to another brother about his previous days fishing experiences. This knowledge might be shared with close blood relatives or might be conveyed in general to others. The results also suggest that the social memory was limited to the days of that same week so the immediate social memory was at most up to 6 days. Sundays represented an opportunity to restart the memory for the next week.
Currently the model uses two-objectives, but
the results suggest that there may be evidence for
other sub-goals in the acquired data set. Future
work would be to expand the hierarchy of goals
for agents. Also, there were gaps in some of the
Pareto fronts, suggesting environmental
constraints may make optimal decision making in
those regions problematic (infeasible). Future
work will be to investigate those areas of the curve
to identify the reasons why. There is also potential
to integrate a virtual reality implementation to
show fish movement dynamics using a: Fish
Visualizer. In addition, we plan to take advantage
of the Social Network capabilities of Cultural
Algorithms in order to attempt to model the
impact that knowledge acquisition and its
subsequent distribution has on strategic
decisionmaking. [