© Copyright JASSS

  JASSS logo ----

Jim Doran (2001)

Intervening to Achieve Co-operative Ecosystem Management: Towards an Agent Based Model

Journal of Artificial Societies and Social Simulation vol. 4, no. 2,

To cite articles published in the Journal of Artificial Societies and Social Simulation, please reference the above information and include paragraph numbers if necessary

Received: 01-Nov-00      Accepted: 01-Feb-01      Published: 31-Mar-01

* Abstract

We propose an advanced agent-based modelling approach to ecosystem management, informed and motivated by consideration of the Fraser River watershed and its management problems. Agent-based modelling is introduced, and a three-stage computer-based research programme is formulated, the focus of which is on how best to intervene to cause stakeholders to co-operate effectively in ecosystem management, and on the objective discovery and comparison of intervention strategies by way of computer experimentation. The agent-based model outlined is technically relatively complex, and several potential difficulties in its detailed development are discussed. Types of ecosystem intervention strategy that might plausibly be discovered or recommended by the model are projected and compared with those currently advocated in the literature.

software agent, agent-based modelling, integrated watershed management, sustainability, Fraser River, intervention strategy.

* Introduction

Achieving stakeholder co-operation[1] for sustainable management of ecosystems throughout the world is widely recognised as a major challenge for the twenty-first century[2]. In this paper, we propose and discuss an approach to meeting this challenge that uses advanced agent-based modelling incorporating representations of stakeholders as well as of the natural resource system.

The idea that agent-based modelling may usefully be applied to improving renewable resource management is not new and much important work has already been reported (e.g Bousquet et al 1994; Bousquet et al 1999; Carpenter et al 1999; Rouchier et al., 2000). However, the particular emphasis in this paper is on discovering how best to intervene in a complex renewable resource management situation, without provision of new resources, in order to bring about enhanced stakeholder co-operation and management effectiveness. Intervention strategies and management regimes for sustainability have been much discussed in the recent research literature (e.g. Christie and White 1997; Healey, 1998; Nielsen and Vedsmand, 1999). The experimental modelling approach to be suggested here potentially enables different strategies to be discovered, assessed and compared. Although the agent-based model to be discussed has not yet been implemented on a computer, and indeed is quite far from implementation, detailed consideration of it gives rise to valuable insights and thought-provoking questions as will appear.

* Agent-based social simulation

Agent-based social simulation is essentially computer simulation, but using relatively complex models that include agents (Doran, 1997; Gilbert and Troitzsch, 1999). In this context, agents are software entities that are "autonomous" loci of decision making. Agents sense, decide and act. Reactive agents choose actions directly by reference to their sensed circumstances. Deliberative agents are those which, for example, reflect upon alternative courses of action, and select one of them for execution. That is, they plan. Adaptive agents change their behaviour in the light of changing circumstances. Social agents communicate and co-operate with other agents.

Agent-based modelling draws upon the rapidly developing technology of agents and multi-agent systems (Jennings and Wooldridge, 1998;Weiss, 1999), a mixture of artificial intelligence and other computer science techniques, notably object-oriented programming. A range of alternative agent software design architectures have been developed, including architectures based upon simple or "fuzzy" rules, upon artificial neural networks, upon AI planning systems, and upon hybrid architectures. Possible relationships between agents include co-operation and negotiation, subordination, and competition. Standardised inter-agent communication languages have been developed notably KQML[3] (see Weiss, 1999, pp 88-92) and FIPA-ACL[4] (see Weiss, 1999, p 592 and http://www.fipa.org/) that are based upon the notion of "speech acts". In agent-based models of components of human societies, agents are often interpreted as in correspondence with individual humans, with small human groups, or with sections or the entirety of organisations.

At its most successful, modelling (not necessarily agent-based modelling) promises major insight into the system modelled, the target system, leading to scientific advance, and promises useful prediction of the target system's behaviour leading to practical policy guidance. Even when, as is all too often the case, these primary objectives are at best only partially achieved there may be other "soft" benefits of a modelling study. For example, a study may encourage the collection of data and the recognition of existing gaps in knowledge. It may enhance communication between researchers and better focus their discussions. It may serve as a useful educational tool. Minimally, a modelling study will illustrate the concepts that inform and guide it. And quite often an attempt at modelling introduces one or more interesting technical problems, which then embark upon a life of their own.

However, the further and particular promise of agent-based modelling is that it enables us to work with computer models that explicitly get to grips with the impact of individual and collective cognition in social systems. It is possible to address the interaction between the real world and the collectively, that is socially, conceived world. Few would seriously deny that the beliefs and goals of individuals often influence and are influenced by the structure and dynamics of human society. Agent-based models with non-trivial agents within them enable such matters to be addressed.

* Ecosystem management

Environmental concerns have become ubiquitous as the global human population increases along with its consumption of natural resources. Instead of poorly-considered local resource exploitation, often competitive and destructive in the longer term, explicit and large-scale management of natural resources -- agriculture areas, forests, fishing stocks -- has been forced upon us as we attempt to solve ever-worsening problems of resource decline and failure. Notions of integrated management, covering all aspects of an ecosystem, and of sustainability, seeking to manage natural resources in ways that secure the future, have become prominent. The precise meaning of sustainable is somewhat elusive. A well-known, and cautiously worded, definition of sustainable development is "development that meets the needs of the present without compromising the ability of future generations to meet their own needs" (WCED 1987:43). Alternatively: "Sustainability means living within the constraints imposed by a finite set of global resources while satisfying the reasonable social and material aspirations of most of the world's citizens." (Healey, 1998a).

Over the past decade a widespread response to perceived past failures has been to reject centralised, "top-down" and "bureaucratic" management, in favour of decentralisation and an allegedly more traditional co-operative and community centred form of management. Since management and its failures impacts peoples' lives in profound ways, social and political reverberations are inevitable. Indeed, certain styles of government and multinational commerce and industry have become major targets of criticism, hostility and even violence. However, it sometimes seems that this trend towards decentralised management is at root no more than an emotional and ideological reaction in the face of a global problem that human society is not currently equipped to address. More objective insights are required.

Integrated Watershed Management

Integrated river watershed management is a category of ecosystem management that is of obvious importance and complexity (see, for example, the studies of the Nile watershed presented in (Abu-Zeid and Biswas, 1996). Potentially conflicting requirements within a watershed are:

Biophysical investigation addresses such matters as water flows, fish population dynamics, types and levels of pollution and their causal processes, and the impact of control procedures. But beyond problems at the biophysical level are human social problems that are sometimes insufficiently considered. Human society in a watershed typically comprises a multiplicity of individual and organisational stakeholders, each pursuing its own interests, with conflicting goal sets, and each with only limited awareness and knowledge of the others. There may well be sharp political and cultural differences. It seems that the consequences of this competitive divergence of interests include resource exhaustion, pollution and social inequality. Put in general terms, the task at all scales in a watershed is to agree and execute a strategic program of natural resource utilisation and conservation which balances sectional interests and which is sustainable and equitable. To achieve sustainable balance is not easy, and typically is impossible without restraint and compromise.

The Fraser River Watershed

An important particular case of watershed management is that of the Fraser River in British Columbia, Canada (but see also, for example, the controversy surrounding the Columbia River in the USA and the Snowy River in Australia). The management task for the Fraser may be viewed on any of three scales: the entire Fraser watershed; the Lower Fraser Basin and separately, the Upper Fraser; or, recursively, any of the sub-watersheds of the Fraser. At each scale there are particular characteristics and problems as well as the problems common to all three scales.

The Fraser River watershed as a whole is large and diverse ranging from conurbation (Greater Vancouver) to near wilderness. Logging is a major industry, and the salmon populations are of major importance but in serious decline. The watershed displays all the aspects and problems mentioned earlier. Over the past decade a major attempt at integrated management of the watershed at this scale has been made, but this has encountered many social and political difficulties and as yet has had limited success (Dorcey, 1997; Marshall, 1998). A particular problem is the complexity of governance. Federal, Provincial, Municipal and First Nations governments are all stakeholders in the watershed. The aboriginal or "First Nation" dimension adds further distinctive complexity as aboriginal culture becomes more influential and, for example, land rights are (re-)negotiated.

The Lower Fraser Basin (Healey, 1999) has experienced a population explosion over the past century, which has resulted in de-forestation, agriculture, industrialisation and conurbation, and rising levels of pollution. There is now a major immigrant population, a majority of Chinese race, whose cultural norms and expectations do not fully align either with those of European or of First Nation origin. This ethnic diversity further complicates the environmental issues described earlier.

The particular problem of the smaller urban sub-watersheds, for example, the Brunette River watershed (Hall, 2000) is that any attempt at integrated management must grapple with a range of powerful stakeholders, including any or all of the four orders of government listed earlier, and major corporations. Achieving a coherent watershed management plan (and financing and then implementing it) becomes a delicate matter of initiating and bringing to fruition negotiation between the representatives of these powerful stakeholders, with local community input where possible.

In sum, study of the Fraser watershed reveals a diversity of stakeholders, the majority organised into decision-making hierarchies, that are currently failing to collaborate to solve both watershed wide and more localised distributed environmental control problems, with the result that key environmental variables are moving out of their acceptable ranges.

* The problem abstracted

Consideration of the Fraser River watershed and its problems makes clear that ecosystem management problems are very often primarily to do with people and organisations and the attitudes that they adopt and which profoundly influence their behaviour. Particular sources of difficulty include the private agendas of the particular stakeholders and their conflicting norms and expectations, stakeholders' lack of information, and failures of inter-stakeholder communication. It is rare that a management regime that is beneficial overall will always be to the benefit of every stakeholder. Very often a management strategy that serves the interests of a majority of the population in the long-term will significantly harm the interests of minorities. Thus there are always issues of compromise and restraint.

It therefore seems that if agent-based modelling is to be used to address watershed management problems, models must address the social and human issues of organisation, conflicting belief systems, partial knowledge, imperfect decision-making, fallible communication, negotiation and the establishment of co-operation. Indeed, it seems that we must focus on social intervention strategies. This insight can be made more structured by distinguishing three levels[5] at which the problem of integrated watershed management can be formulated:

As indicated above, the Fraser River experience (Dorcey, 1997) makes clear that the level I problem, although very difficult in practice (and the problem containing the "hard science"), is not that of the greatest importance. Nor indeed is it the level II social and management structure design problem where the major difficulty lies. The key task is the "people problem" at level III, the task of intervening to reconcile conflicting interests and to engender co-operation. Intervention policies are needed which move a discordant society that is not even fully aware that it has a problem of achieving integrated management for sustainability, to one that actually co-ordinates to solve the problem. Thus the emphasis shifts towards communication, intervention and persuasion.

It is important to make clear that our focus is on intervention without provision of any significant level of new resources. That is, we consider the task of persuading stakeholders to co-operate effectively without the deployment of significant financial aid and without changes to the physical infrastructure or enhanced communications. Intervention on these terms is very common, for example by a government, NGO or academic team, seeking to bring about an improvement in a natural resource system which is under some kind of threat. At present the methods and principles by which such intervention should be made are a matter of controversy and assumption. There is a need for deeper and more objective insights. Agent-based modelling with deliberative agents is of potential value just because it can plausibly extend computer modelling into this difficult area.

* Using agent-based modelling

To gain insight into alternative intervention possibilities we examine the following three-stage research programme:
  1. Design an agent-based model of the intervention scenario that is sufficiently precisely specified to be implemented on a computer, and implement it.
  2. Experiment with the model to explore possible social interventions and to discover those that (within the model) appear to be effective
  3. Test the discovered social interventions in practical contexts or otherwise.

We discuss each of these three stages in turn, starting by formulating an outline model for a watershed ecosystem such as that of the Fraser River. The focus is upon human society in the watershed, and on the intervention strategies by which society may be acted upon.

Stage 1: the Agent-Based Model

Some preliminary remarks are needed. Firstly, and crucially, the model is described subject to the constraint that it must be programmable for a computer. Wherever words are used in the model specification such as "goal", "belief", "behaviour", "plan", "cognition", "action", "message", "society", they should be understood to have the meanings regularly given to them in the context of software agent design and implementation (e.g. inWeiss, 1999), not the meanings of their normal everyday use.

Secondly, the previous remark notwithstanding, there is no attempt to state the model in any rigorously formal way e.g. within a mathematical logic. To do so at this stage would merely deflect from essentials.

Thirdly, the model is highly abstract. The aim is achieve new and more precise understanding of the typical intervention scenario, rather than to draw specific conclusions about, for example, intervention in the Fraser watershed.

Finally, the model is individual-based, with a loose correspondence between people and software agents.

The model takes as its start point the assumption that there exist centralised hierarchies of agents, with overlapping spheres of action, which are competing in an initially resource rich environment. In a little more detail:
  1. There is a large society of software agents (>1,000). The society is dynamic in the sense that it is running on a computer with ongoing actions by agents and interactions between agents. The agents are heterogeneous in the sense that different agents have different behaviours and beliefs, and different types and degrees of cognitive ability.
  2. The agents are organised into competing hierarchies. At the top of each hierarchy is an agent (or are agents) whose goals reflect the resource interests of the hierarchy as a whole. Agents lower in the hierarchy normally act loosely in accordance with more limited goals set by their immediate superiors.
  3. The society is set in a simulated natural environment incorporating an abstract resource management task that impacts all agents. Solutions may be assumed typically to require extensive agent co-operation. The hierarchies have overlapping spheres of activity.
  4. Intervention is possible in the sense that the "experimenter" may at any times send messages of specified types to some or all of the agents within the society.

To take this model specification towards an actual computer implementation, we comment briefly on each of the four components in turn.
The agents

As has been emphasised, agents are software entities. There exists a range of available agent designs that might be adopted for a model such as this, but none entirely suitable. The simplest agent architectures use a small number of condition-action rules to determine action. However, in this context it is hard to avoid the conclusion that agents should include beliefs, goals and, to some degree, internal models of their "social" context and the environment. They need the ability to act, perhaps via the creation and execution of plans, and the ability to send and receive messages. One way to express these requirements is to say that agents must have an operational Belief, Desires and Intentions (BDI) architecture (see Weiss, 1999, p 585), including the ability to process incoming and outgoing speech-acts. Such agent architectures have indeed been developed, but only in experimental research contexts and with cognitive abilities still very limited compared with human beings. These architectures have rarely if ever been used for agent-based social simulation.
The agent hierarchies

There are many ways in which a set of software agents may be engineered to function as an organised, centralised hierarchy (see, for example, Prietula, Carley, and Gasser 1998). Perhaps the simplest method is to include within each agent specific condition-action rules that straightforwardly lead it to play its required role within the hierarchy. This role will include responding appropriately to messages from subordinates, and the completion of tasks delegated to it from superiors (which may involve sub-delegation). The rules may also support messages being exchanged with other agents at the same level in the hierarchy. A major limitation of this rule-based agent architecture is that the agents are not themselves able to vary or withdraw from their own role, so that the hierarchy is inflexible.

A different and more "realistic" approach is to design agents in terms of the (self-interested) goals that they contain and seek to achieve. An agent's role within the hierarchy then flows from its goals and the plans that it creates and executes to achieve them. The agent decides to undertake the appropriate role tasks for the rewards that it believes will accrue to it by doing so. Clearly this requires that the hierarchy functions to deliver rewards to agents, or, at least, that agents believe that it does. This approach is more sophisticated, and the hierarchies formed more flexible, but is also much more complex because it requires that agents be designed and implemented to adopt, choose between, delete, and work to goals as a primary activity (c.f. the BDI architecture mentioned earlier). We assume that the agents in such a hierarchy will be heterogeneous in various ways, notably in the autonomous goals that they possess, the beliefs about the environment and the hierarchy itself that they possess, and their actual ability to plan and execute plans including plans involving others.

Another key aspect of any multi-agent system is inter-agent communication. In this context ad hoc communication is likely to be sufficient, but a standardised inter-agent communication language is also a candidate.
The environment sub-model

The natural environment sub-model must capture the need for long-term co-operation to achieve sustainability, with failure to co-operate by a significant proportion of agents leading to collective disadvantage (see Hardin's "Tragedy of the Commons", Hardin 1968). Co-operation implies short-term restraint by agents, rather than "greed", and both local and global action by agents. The sub-model should also permit the issue of equality or inequality of agent access to resources to be addressed, since it is commonly held that sustainability implies not only the preservation of the resource base, but also the achievement of fair access to it.

In a typical watershed there are many variables and stakeholders of relevance. Potentially significant variables include: input flows to tributaries, irrigation requirement and take, fish stock levels and catch, levels of use of agricultural chemicals, pollution levels, dam flow settings, hydroelectric power generation levels, dam & lake recreation requirements. Stakeholders include local and national government, farmers, fisheries and river conservationists, hydropower companies, and recreation organisations.

A suitable form for the environment sub-model might thus be set of variables and parameters divided into (overlapping) subsets as follows:

together with a set of non-linear recurrence relations, which specifies the interconnections between the variables. Spatial distribution would be implicit in the model structure. The sphere of activity of a particular agent (or agent hierarchy) would then be that subset of all the variables and action parameters that is accessible to it. Limited spheres of activity imply a need for inter-agent co-operation.

Put in the most general terms, the intervention task is then to influence the agents so that certain selected variables are kept within specified limits, in the face of fluctuation in independent environmental variables. The agents can refer to sensed variables and can adjust those variables that they collectively have within their control. Just which variables are to be maintained within which limits (so achieving sustainability and equity between agents) is taken as inherent to the intervention task itself. Specific ecosystems, for example the Fraser watershed, would give rise to specific environment sub-models, with a consequent requirement for model validation. Fortunately, the search for effective interventions can, and arguably should, be made in terms of typical environment models.
The possible interventions

As stated earlier (see 4.4), we are restricting ourselves to interventions that involve no significant new resources. This means that interventions take the form only of messages sent by the experimenter to agents in the model. Interventions that directly impact the natural environment sub-model are excluded. The intervention language should precisely specify the syntax and semantics of the messages the experimenter may send to agents. As in the case of inter-agent communication, the language could be ad hoc or based upon an established language such as KQML or FIPA-ACL. Messages might be limited to assertions and questions, or include such content as suggested goals and plans. Whatever intervention language is used, agents must be able to receive, assimilate and act upon -- or decline to act upon -- the allowable messages and their content. This may well involve complex issues of belief and goal adjustment, depending upon the complexity of the agents themselves.

Stage 2: Discovering Intervention Strategies

Assuming that the model (comprising agent hierarchies, environment sub-model, and intervention set) has been formulated in detail and implemented on a suitable computer, the next task is experimentally to study possible interventions in search of those leading to sustainability and social justice. Essentially, an intervention is a single message sent to a single agent or "broadcast" to a particular set of agents. Within a precise model, the set of all possible combinations of interventions is well defined, assuming that the timing of the messages is specified in some suitable way. The performance of a set of interventions within the model may be assessed by their effectiveness as measured in the natural environment sub-model.

The discovery task involves repeatedly setting the model running, and then observing the effect of particular combinations of messages sent to the agents. Often there will be a degree of non-determinism, so that the same combination of intervention messages may yield different outcomes on different occasions even for an identical initial set-up of the model. Note that the "null" case, where the model is allowed to run with no messages sent, may itself produce highly complex behaviour and requires special investigation.

The set of all possible interventions is clearly extremely large. Indeed it is easily made infinite. The task of searching for effective combinations therefore seems close to intractable. However, much can be done automatically, at least in principle. Many relevant search techniques are available in the artificial intelligence repertoire, including varieties of hill-climbing search and of genetic algorithms (Russell and Norvig, 1995). Furthermore it is appropriate to try to work in terms of intervention strategies, that is, sets of interventions organised and targeted in some particular way, and then to reformulate the task as that of searching for effective intervention strategies.

Stage 3: Testing Discovered Strategies

Suppose that we find an effective intervention strategy as judged by its performance within the model, that is, by its impact upon the agent society and the society's interaction with the simulated natural environment. How can the discovered strategy be put to the test and validated outside the model? There seem to be two main possibilities:

The first of these possibilities is clearly a major undertaking with substantial financial, social and political implications and risks. The second is much more immediately possible, but a degree of objectivity is lost since the assessment criteria become more subject to influence by current theoretical preferences and assumptions (see section Comparisons).

A third but more remote possibility is to use the agent model as the basis of an artificial role-play scenario in which stakeholders "play themselves" (compare Bousquet et al., 1999). It seems doubtful, however, whether stakeholders playing these roles could respond to interventions as they would in real life, particularly were any of the more "Machiavellian" strategies involved (see later sections Intervention strategies and Comparisons).

EOS Revisited: a Simple Illustrative Example

Insight into the model outlined and into possible experiments using it may be obtained by taking as an illustrative example the artificial society developed as part of the EOS project (Palmer and Doran, 1993; Doran et al., 1994; Doran and Palmer, 1995). In this project an agent-based model of social processes in a natural environment was designed and implemented in the computer language Prolog. However, the work was not directed to resource management issues. Rather it was targeted at the emergence of centralised hierarchies of agents in the context of a simulated spatial environment in which self-renewing resources had to be located, "harvested" and "consumed" by agents if they were to support their "energy" expenditure and survive. Agents were deliberative and relatively complex. Specifically, fluctuating groups of agents negotiated and collectively executed multi-agent harvesting plans, and those agents that over time were most successful in the negotiations came to be recognised by the agents themselves as group leaders or super-leaders.

A substantial programme of experiments demonstrated that the trajectory over time followed by an EOS agent society typically led to a society composed of multiple agent hierarchies competing to harvest the available resources. Each particular hierarchy was relatively efficient at co-ordinating its members for the execution of its own joint harvesting plans but, of course, since there was no communication or co-operation between hierarchies, substantial inefficiencies arose as hierarchies (accidentally) disrupted one another's plans.

In the EOS research there was no consideration of outside (i.e. experimenter) intervention in the sense that we are discussing it here. So let us now briefly consider how the EOS environmental problem might be made of more relevance to our present line of investigation, and what types of intervention might then be applied and what their effects might be.
A modified EOS environmental scenario

In the standard EOS spatial environment, individual resources were normally set to self-renew regularly, providing ample "energy" in total to "carry" the agent population. However, periods of relative resource scarcity could be specified in order to determine the effects they would have upon the emergence of hierarchies. Assume the resource dynamics to be set to an extreme so that:
  1. resources last indefinitely (when not harvested) but do NOT renew once harvested, and
  2. the agents are (slightly) altered so that when they have the opportunity to do so, they tend to consume more resource than they actually need for survival, that is, to "waste" energy.

Note that (i) implies that all agents must ultimately "die" once the initial supply of resources is totally harvested. It seems clear that maximal survival of the agent society requires all agents to have balanced access to unharvested resources, no resources to be left permanently unharvested, and no wastage. However, without intervention, this is not what will happen in the modified EOS scenario because:
  1. by assumption the agents tend to over-consume when able to do so, and
  2. as the agent hierarchies do not co-operate, inefficiency in harvesting resources leads to wasted "energy" (for example, unsuccessful journeys) and therefore unnecessary resource consumption.

    Furthermore, it is quite possible that:

  3. because the competing hierarchies do not share information, some agents may "die" unnecessarily early because they are unaware that unharvested resources still exist and are still available to them.

Although this modified EOS scenario is crude and unrealistic, notice that it it does capture a requirement for the agents to co-operate to avoid waste and over-consumption, and to achieve balanced access to the available resources.
Possible EOS interventions

Our interest is in interventions that have some degree of realism, avoiding, for example, massive and arbitrary interventions that impact the agent society in an unconstrained way. Recall that agents in a typical multi-agent system such as EOS have only partial and possibly inaccurate knowledge of the system as a whole and of the simulated environment. Suppose that interventions are restricted to those that supply information to agents, and that the agents always accept this information. Of necessity, information provided could only be of a type that the agents can represent internally and work with. Furthermore, it is implied that any previous but contradictory information held by an agent is deleted. What now can be done?

Perhaps the simplest option is provide all agents with accurate and continuously updated information specifying the locations of all unharvested resources. This deals with problem (c) of the preceding section, because now all agents know all there is to be known about unharvested resources and their locations. But it does not deal with problem (b), the inefficiencies caused by lack of co-ordination. One method of dealing with problem (b) would be regularly to mislead certain of the agents (primarily the heads of hierarchies), by supplying them with false information about resource locations. Essentially the experimenter would take detailed control of which agent harvested what at each time, in the interests of overall efficiency, by systematic distortion of the agents' knowledge of resource locations (their "cognitive maps"). One might describe this as agents being sent on "wild goose chases" as necessary. This intervention strategy might also be used to prevent individual wasteful consumption (problem (a) above), because agents could be prevented from ever knowing that they are in a position to consume further.

A more radical intervention strategy might aim to merge the existing agent hierarchies into one, thereby ensuring a greater degree of co-ordination and so addressing problem (b) above. Hierarchies exist, within the EOS model, solely as patterns of belief within agents. Each agent has beliefs about which other agents are its superiors and inferiors, and acts accordingly. It follows that sending messages to agents that change these beliefs can alter hierarchies. For example, if the top agent of one hierarchy were to be led to believe itself the follower of the top agent of another, and that other agent were led to believe itself the leader of the first, with appropriate changes to secondary beliefs in both agents, then the two hierarchies in question would be merged into one.

A difficulty with these various intervention strategies is the highly unrealistic assumption that the agents will always accept the information sent to them. The assumption implies, for example, that agents will accept X, not X, X and not X in rapid succession without question. The interventions proposed also assume that the experimenter has access to all information and can communicate with all agents at all times. This is plausible enough in the actual experimental situation, but not in reality. Clearly these assumptions need to be weakened and moved closer to reality. Nevertheless this EOS example does illustrate some of the issues that arise when considering intervention strategies in detail.

The Technical Problems

The foregoing description of the suggested research programme, and of the modified EOS scenario, touched upon certain theoretical and practical difficulties that will inevitably be encountered in carrying it through. We now briefly focus upon these difficulties.

Specifying the model -- The model specification presented earlier (5.7 et seq) is certainly not fully precise, and is in many ways arbitrary. To develop the specification to one of full precision is certainly possible (in the same way that the original EOS specification was fully precise), but it will always be difficult to justify one model variant rather than another. For example, what type of internal representation (if any) of the agent hierarchy of which it is a member should an agent maintain and use? What range of inter-agent communications should be implemented? Furthermore, limitations of current agent technology mean that the complexity and cognitive powers of the software agents will be very limited, with all that that potentially implies for realism.

Searching the space of interventions -- As noted earlier, searching the space of all possible interventions is a huge combinatorial task. Fortunately, automatic heuristic search techniques are available. Working in terms of intervention strategies is attractive, but begs the question of the precise definition of an "intervention strategy". Even if intervention strategies can be precisely formulated in terms of a specified intervention language (and this will not be easy), their effectiveness might depend greatly upon details of the agents, of the agent hierarchies, and of the resource problem. This would imply that there is no simple and general "map" of the exploration space to discover. At worst the impact of particular interventions might prove to be entirely unpredictable in all circumstances, but this would be counter-intuitive.

Testing the interventions -- This is, essentially, the issue of model validation. As noted earlier, it is a potential test of intervention strategies derived from the model to ask if they can be systematically related to those in current practice and in the relevant literature in an insightful way (see sections Intervention Strategies and Comparisons). Actually to deploy the suggested intervention strategies in practice and observe their real outcomes would be a much more demanding, expensive, time-consuming, and risky test procedure, but ultimately more persuasive.

Many of the difficulties noted here are instances of quite well recognised general problems with agent-based social simulation as discussed in, for example, (Doran, 2000). Unfortunately these practical difficulties, although substantial, cannot be avoided. There is no "free lunch".

* What intervention strategy might be found?

The promise of working with a computer-based model is that some pre-conceptions may be avoided and objective discoveries made. The implication is that to try to second-guess the model, without actually running it, is precisely to throw away the advantage that modelling offers. On the other hand, it has often been remarked that it can be thought provoking, and can offer valuable insight, merely to formulate social problems and processes in computational multi-agent terms. What, therefore, can be said now about strategies for intervention in the proposed watershed model, assuming that these are in pursuit of sustainability and social justice?

Classes of Intervention Strategy

Keeping always in mind that we are here dealing with organisations of software agents on a computer, the following broad classes of intervention strategy seem intuitive given the abstract model as specified. Recall that success is a matter of finding collective agent behaviour that solves "environmental" problems of the type discussed earlier as captured in the natural environment sub-model. For each class of intervention strategy a few comments are added from the agent technology perspective.

S1 -- Command The simplest intervention strategy is for the experimenter to decide what needs to be done, and then to take direct control of all agents (or perhaps just of "heads of hierarchies") giving them detailed and suitable instructions. However, whether agents are straightforwardly rule based, or whether they are more deliberative and goal driven, with some designed degree of autonomy, this strategy will not often work. Agents will either be incapable of receiving their instructions, or will receive them but not necessarily comply with them.

S2 -- Fact Provision Assuming that the environmental problem posed has a mutually beneficial solution (arguably an unrealistic assumption), then an appropriate strategy seem to be merely to inform agents of the facts of their situation and to leave it to them to work out what needs to be done. Bearing in mind that agents in a multi-agent system typically have only partial knowledge of the system as a whole, initial questioning might be needed to establish what the agents already know. Put otherwise, the agents, rendered more knowledgeable by the intervention, are in a position to work out a solution to the environmental problem for themselves. Of course, this presumes that the agents are, in an appropriate sense, willing to accept the information supplied to them and have the cognitive abilities to make use of it (see the earlier discussion of BDI agents and of the EOS scenario).

S3 -- Misrepresentation Alongside the "fact provision" intervention strategy is the possibility of sending messages to the agents in the society that provide them with plausible but partially inaccurate or even entirely false information. The cumulative effect of this unreliable information may nevertheless be to bring about the desired collective action. At first sight this may seem paradoxical, but it is not, as was illustrated in the EOS example (see 5.31 et seq). For example, the effect of informing all agents that environmental dangers are much greater and more immediate than they actually are, may be to steer their decision making in the direction of a solution to the actual environmental problem (cf. Doran, 1998). This raises a technical issue in agent design. How and under what circumstances should an agent come to "believe" information supplied to it? The answer is not at all straightforward in general, and in this particular context there seems to be a need for a practical theory of "plausible" lying.

S4 -- Communication Enhancement This strategy is straightforwardly to encourage more inter-agent communication. This might be done by, for example, merely making agents newly aware of possible communication channels, or by "persuading" them to favour goal achievement strategies involving communication with others. Of course, the strategy will depend on the relevant enabling "cognitive" mechanisms having been included within the agents themselves. The underlying target of the strategy might be to enhance the spread of (mis-)information amongst the agent community. Recall that we have chosen to exclude from consideration actual "physical" enhancements of the simulated natural environment, such as new roads.

S5 -- Focus Adjustment At any particular moment any complex agent has the processing capability to work on only a limited subset of the cognitive tasks that it in principle might be working on. That is, it must have a focus of attention. In particular, if an agent has a number of goals pending, it must select at most a few of those goals to pursue at any time, with consequent selections of sensory information considered and knowledge deployed. This is true also of agent organisations. It follows that an agent society may fail to address certain issues of importance to it, not because it lacks awareness, knowledge, or the requisite cognitive ability, but simply because its focus of attention is elsewhere. This opens the possibility of a different type of intervention strategy, designed to redirect focus of attention. Just how such a strategy would work is unclear, but it must surely interact with the focus of attention processes present within the agents and the agent organisations.

S6 -- Destabilisation The final intervention strategy that merits mention is that of sending messages to the agents, the effect of which is to de-stabilise the hierarchy system with the hope that it will later self-organise more effectively. This might be achieved by, for example, prompting agents to engage in unproductive activities which in turn leads them to withdraw from current organisational roles. Bizarre though it might at first seem, this strategy makes sense if there is reason to believe that the agents system is locked into an inferior stable state and needs a significant "jolt" to change to something better. It implies that emergence of macro structure and process from the micro level, and how exactly that proceeds, is an important consideration.

These six strategies are all meaningful in software agent terms, although some of them presume agents of a "cognitive" complexity that at present we barely know how to build on a computer. The strategies are not, of course, wholly independent. They might well be used in combination. For example, "fact provision" may well be coupled with "enhancing communication", and "misrepresentation" coupled with "destabilisation".

Centralisation and Decentralisation

Independently of the actual choice of intervention strategy, the execution of any particular strategy may often be divided into two steps. These steps are first the collection of information about the agent hierarchies and the design, if appropriate, of a revised organisation for them, and then secondly the actual process of intervention to bring about the chosen organisation and to set it in operation. The first of these steps loosely corresponds to the level II problem formulation as defined earlier, and the second to the level III problem formulation. Choosing an organisational structure for agents prompts consideration of centralisation and decentralisation and criteria for their effectiveness. Accordingly we now discuss the impact of hierarchies on collective task performance.

From a computational perspective, centralised hierarchical organisation potentially enables one agent (or a small number of agents) to take and execute decisions in the light of a single abstract overview of the task scenario, deploying appropriate abstract knowledge. This may be of crucial benefit to efficient collective task performance. But this benefit will not arise if
  1. the task simply does not require such an overview for its performance (for example, if local problems do not inter-relate), or if
  2. no adequate abstraction scheme exists, or if
  3. the flow of information up the hierarchy and of action down the hierarchy frequently fails or is too slow.

Point (ii) requires some elaboration. If there is no effective method available to the hierarchy of abstracting and summarising the task and the knowledge required to manage it, then any attempt at centralised overview and management is doomed to failure or, at best, overload in cognitive processing.

These insights, although they originate in the study of artificial agent societies, apply equally to human societies. The objective of decentralisation should thus be to recognise and address these difficulties with centralisation, insofar as they exist in the particular case, by striking a balance between the actual need for and feasibility of central overviews and the delays and errors inherent in any substantial attempt to create them. Thus from this perspective intervention needs to establish what degree of decentralisation is most effective and then to achieve it.

* Current strategies for ecosystem intervention and management

Most practical strategies for intervention in ecosystem management situations have in common that they seek to achieve some combination of: Normally there is no attempt to dismantle or significantly alter existing organisations.

A number of ecosystem management strategies are prominent in the environmental resource management literature. They reflect a current emphasis on decentralised and co-operative ecosystem management, and include:

These different strategies may be summarised as follows:

By contrast:

Notice that these existing strategies typically put the emphasis on the management regime to be achieved, rather than on the intervention process by which it is to be achieved. The intervention process itself is typically assumed to be a matter of contact, discussion and persuasion.

A further possibility is merely to act to improve the resource sub-system (for example by cutting a new canal, or introducing a new and better yielding crop variety), without seeking to change, or even understand, the associated human socio-cultural system. However, there is much evidence to suggest that this approach can easily prove counter-productive (Ostrom, 1995) and, as stated earlier, it is not considered here.

These various management strategies, although all in practical use, are not all of the same type, are not precisely defined, and are not exhaustive or well differentiated. Their limits of applicability are little considered in the literature except in the vaguest terms. The question naturally arises, therefore, which of these strategies should be used in which circumstances? And what other possible ecosystem management strategies are there? It is precisely such questions that the agent-based model specified here might go some way towards answering.

* Comparisons

Let us now briefly compare the strategies described in the preceding section with those obtained by "second-guessing" the agent-based model (see Classes of Intervention Strategies). It is apparent that all practical intervention strategies involve elements of knowledge provision, implicitly or explicitly, and seek to enhance inter-stakeholder communication in particular ways. This is particularly clear in Maximal Stakeholder Involvement. Thus the strategies S2--Fact Provision and S4--Communication Enhancement, derived from the model, are prominent in actual practice.

The Co-Management strategy seems best analysed by reference to the issues around centralisation and decentralisation discussed earlier (see section Centralisation and decentralisation). The aim of this strategy is to replace centralised hierarchies with something more decentralised. The implication is that fully effective abstract overviews are not feasible. The Community Management strategy goes further and puts even greater emphasis upon distributed management and organisation, and rather little on hierarchies, and therefore promotes multi-agent system co-operation regimes that do not construct any unified overview. But it is clear from the brief analysis of the section Centralisation and decentralisation that decentralisation is by no means always appropriate, depending on the facts of the situation. Unfortunately, the arguments in the literature for community management or co-management and decentralisation are rarely formulated in precise mathematical or computational terms and rarely, indeed, seem to allow the possibility that decentralisation might not be appropriate. Thus there is clearly work to be done here to reconcile these two conflicting conceptual repertoires.

The Adaptive Management strategy is a bird of a quite different feather. It is less obviously than the others an intervention strategy, and has relatively more to do with what intervention is intended to achieve. Adaptive management seems no more than common sense when there is either initial ignorance of the details of the management task or if the task is changing. Why might it not be employed? The answer perhaps lies in the technical agent notion of commitment. To minimise computational load, it is always attractive for an agent (software, human or organisational) to commit to courses of action and not to reconsider them other than in exceptional circumstances. Of course, commitments can go seriously wrong if reluctance to withdraw them implies a failure to adapt to changing circumstances that is inappropriate or excessive. Hence, perhaps, the need to correct over-commitment by promulgating a notion of flexibility in management. In this strategy, therefore, there are elements not only of S2--Fact Provision and S4--Communication Enhancement, but also of S5--Focus Adjustment and perhaps even S3--Misrepresentation.

Finally, certain strategies suggested by consideration of the agent-based model rarely appear in the natural resource management literature. Recall that in the discussion of the model we suggested that what might be called "Machiavellian" strategies, those that rely upon misrepresentation and creating a misbelief system (S3--Misrepresentation), could be effective, with collectively beneficial outcomes. This possibility is rarely mentioned (but see Rouchier et al, forthcoming) although ideologies and their effectiveness are prominent in the sociological literature. Of course, many would object on principle to the notion that one might proceed by means of a deliberate policy of misrepresentation, however desirable the anticipated ends. Similar objections would be made to the idea of deliberate destabilisation (S6--Destabilisation) of the target society. But one need not look far to find real world examples.

* Conclusions

We have considered the Fraser River watershed, and examined in some detail a possible agent-based model of an ecosystem, including stakeholders. We have discussed some of the intervention strategies the model suggests and their relationship to established (or, at least, advocated) management procedures. The following conclusions may be drawn:
  1. Study of the Fraser River watershed and its resource management problems indicates that emphasis should be placed on an agent-based model that supports the experimental discovery of intervention strategies to achieve co-operative management by watershed stakeholders, rather than on a more conventional agent-based model.
  2. More detailed analysis suggests that the types of intervention strategy that such a model would bring forward are indeed related to, but are by no means the same as, those discussed in the eco-system management literature. The implication is that there are significant discoveries of practical relevance to be made.
  3. The model proposed here goes well beyond existing work in agent-based social simulation. However, there are a number of technical difficulties to be overcome before it can be taken to implementation. Prominent amongst these difficulties are the design and implementation of rather more complex and "intelligent" agents than those currently available, and the discovery of ways to search the near infinite space of possible interventions to locate those effective in particular types of environmental context.
  4. Overall the suggested programme of research is viable but challenging, and potentially requires very substantial effort over a number of years. Perhaps the most critical question is whether a successful balance can be struck in the model between realism and tractability. However, the results of such a study could well be of very real importance and would be obtained much more safely and cheaply than by than by experiments with real ecosystems.

The central argument of this paper is that here is an opportunity to assess possible ecosystem intervention strategies, and perhaps to generate new and effective ones, from a computer model alone. We might even hope to discover new and useful concepts in terms of which to discuss and formulate ecosystem intervention. A prospect is thus opened up that, although technically challenging, is exciting and goes well beyond the present limited use of agent-based models.

* Acknowledgements

I gratefully acknowledge hospitality in February and March 2000 provided by the Institute for Resources and Environment of the University of British Columbia, Canada, and its Director, Professor Les M. Lavkulich, which gave me the opportunity to learn something of the Fraser River watershed and its associated management problems.

* Notes

1 Throughout this paper "co-operation" is used in a general sense akin to "co-ordination" (see Doran et al., 1997) without reference to the more specialised concept of a "co-operative" and its particular connotations.

2 E.g. "This meeting will be remembered as the moment when governments abandoned the promise of global co-operation to protect planet Earth." Greenpeace final statement after the failure of the Hague climate talks (UNFCCC COP6 ), November 2000. http://greenpeace.org/~climate/climatecountdown/

3 Knowledge Query and Manipulation Language

4 Foundation for Intelligent Physical Agents -- Agent Communication Language

5 These three levels should not be confused with the three possible scales of consideration of the Fraser River watershed described earlier.

* References

ABU-ZEID M.A., and Biswas A.K. (1996) River Basin Planning and Management. Oxford University Press.

BOUSQUET F., Cambier, C., Mullon C., Morand P., Quensiere J. (1994) Simulating Fishermen's Society. In: N. Gilbert & J. Doran (eds.) Simulating Societies. UCL Press. Pp. 143-164.

BOUSQUET F., Barreteau O., Le Page C., Mullon C. and Weber J. (1999) An Environmental Modelling Approach: the Use of Multi-Agent Simulations. In: Advances in Environmental and Ecological Modelling (Eds. F. Blasco and A. Weill). Paris, Elsevier. Pp. 113-122.

CARPENTER S., Brock W., and Ghanson P. (1999) Ecological and Social Dynamics in Simple Models of Ecosystem management. Conservation Ecology 3(2): 4. http://www.consecol.org/vol3/iss2/art4

CHRISTIE P. and White A.T. (1997) Trends in Development of Coastal Area Management in Tropical Countries: From Central to Community Orientation. Coastal Management, 25,155-181.

DORAN J. E. (1997). From Computer Simulation to Artificial Societies. Transactions of the Society for Computer Simulation International, 14, 69-78.

DORAN J. E. (1998) Simulating Collective Misbelief. Journal of Artificial Societies and Social Simulation, Vol 1(1) https://www.jasss.org/1/1/3.html

DORAN J. E. (2000) Hard Problems in the Use of Agent-Based Modelling. Social Science Methodology in the New Millennium. Proceedings of the Fifth International Conference on Logic and Methodology, October 3rd-6th, Cologne, Germany. Published as CD-ROM.

DORAN J. E. Palmer M, Gilbert G N, and Mellars P (1994) The EOS Project: modelling Upper Palaeolithic Social Change. In Simulating Societies: the Computer Simulation of Social Phenomena. (eds. N Gilbert and J Doran) UCL Press: London. Pp. 195-221.

DORAN J. E. and Palmer M. (1995) The EOS Project: Integrating Two Models of Palaeolithic Social Change. In: N. Gilbert & R. Conte (eds.). Artificial Societies London: UCL Press. Pp. 103-125.

DORAN J E, Franklin S, Jennings N R and Norman T J. (1997) On Cooperation in Multi-Agent Systems. The Knowledge Engineering Review, 12(3), 309-314.

DORCEY A. H. J. (1997) Collaborating Towards Sustainability Together: The Fraser Basin Management Board and Program. In: D. Shrubsole & B. Mitchell (eds.). Practising Sustainable Water Management. Canadian Water Resources Association. Available at http://www.interchg.ubc.ca/dorcey/chcwra/fccwra.html

GILBERT G. N. and Troitzsch K. G. (1999) Simulation for the Social Scientist. Open University Press.

HALL K. (2000) Brunette Basin Watershed Plan. Policy and Planning Department, Greater Vancouver Regional District. Vancouver, British Columbia.

HARDIN G. (1968) The Tragedy of the Commons. Science, 162, 1243-1248.

HEALEY M. (1998) Paradigms, Policies, and Prognostications about the Management of Watershed Ecosystems In: R. J. Naiman & R. E. Bilby (eds.) River Ecology and Management. Springer. Pp. 662-682.

HEALEY M. (1998a) Barriers and Bridges to Sustainability in the Fraser Basin. Invited address to the State of the Basin conference, November 20 & 21st, 1998. Vancouver.

HEALEY M. (ed.) (1999) Seeking Sustainability in the Lower Fraser Basin: Issues and Choices. Institute for Resources and the Environment, Westwater Research Centre, University of British Columbia, Vancouver.

JENNINGS N. R. and Wooldridge M. J. eds., (1998) Agent Technology: Foundations Applications and Markets. Springer: Berlin.

JENTOFT,S., McCay B., and Wilson D.C. (1998) Social Theory and Fisheries Co-Management, Marine Policy. Vol. 22, No. 4-5, 423-436.

LANSING J. S. (2000) Anti-Chaos, Common Property, and the Emergence of Cooperation. In: T. A. Kohler & G. J. Gummerman (eds.). Dynamics in Human and Primate Societies. Oxford University Press. Pp. 207-223.

LANSING J S, Kremer J N, and Smuts B.B. (1998). System-Dependent Selection, Ecological Feedback and the Emergence of Functional Structure in Ecosystems. Journal of Theoretical Biology 192, 377-391.

MARSHALL D. (1998) Watershed management in British Columbia: The Fraser Basin Experience. Environments, Vol. 25, No. 2/3, 64-79.

NIELSEN J. R. and Vedsmand T. (1999) User Participation and Institutional Change in Fisheries Management: a Viable Alternative to the Failures of 'Top-Down' Driven Control? Ocean and Coastal Management, 42, 19-37.

OSTROM E. (1990) Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press: Cambridge, UK.

OSTROM E. (1995) Constituting Social Capital and Collective Action. In: R. O. Kehane & E. Ostrom (eds.). Local Commons and Global Interdependence: Heterogeneity and Cooperation in Two Domains. Sage Publications: London. Pp. 125-160.

PALMER M. and Doran J E (1993) Contrasting Models of Upper Palaeolithic Social Dynamics: A Distributed Artificial Intelligence Approach. In Computing the Past -- Proceedings of Computer Applications and Quantitative Methods in Archaeology Conference 1992 (CAA 92) (eds. J Andresen, T Madsen, I Scollar). Aarhus University Press. pp 251-262.

PRIETULA M.J., Carley K.M., and Gasser L. (1998) Simulating Organizations. AAAI & MIT Press; Menlo Park, CA.

ROUCHIER J., Bousquet F., Le Page C., and Bonnefoy J.-L. (Forthcoming). Multi-Agent Modelling and Renewable Resource Issues: the Relevance of Shared Representations for Interacting Agents. Proceedings of the Second Workshop on Multi-Agent Based Simulation (MABS 2000), Boston, 8-9th July, 2000. Springer LNAI series 1979, 181-198, eds. S. Moss & P. Davidsson.

RUSSELL S. J. and Norvig P. (1995) Artificial Intelligence: a Modern Approach. Prentice Hall.

WALTERS C. J. (1986) Adaptive Management of Renewable Resources. Macmillan, New York, USA.

WCED (1987) Our Common Future. Oxford: Oxford University Press.

WEISS G. (ed.) (1999) Multiagent Systems. The MIT Press: Cambridge, Massachusetts and London, England.


ButtonReturn to Contents of this issue

© Copyright Journal of Artificial Societies and Social Simulation, 2001