استراتژی تطبیقی ​​مبتنی بر اطلاعات برای بهره برداری از منابع در سناریوهای رقابتی

Given an exploitation problem, in which a number of agents compete for a limited renewable resource, the optimal harvesting strategy depends on the ratio between resource availability and exploitation effort. For scarce resource a purely competitive, greedy strategy outperforms a more collaborative approach based on the Collective Intelligence, while for more abundant resource the opposite holds. The rationale for this behaviour lies in the amount of information each strategy is able to provide and a combined strategy is possible according to which agents choose dynamically the most informative strategy according to a minimum entropy criterion. This approach, which provides best performance for both under and over-exploited scenarios, can be used to monitor the resource status for management purposes and is effective in both centralised and decentralised decision making.

Several natural phenomena can be cast within the framework of competition: living beings compete for survival, species compete for niches, humans compete for financial resources and, adventuring ourselves into more questionable conjectures, cultures compete for supremacy, ideas compete for our attention [1], and natural selection is even suggested to operate in the cosmos [2]. These ideas are linked to different versions of Darwinism: best competitors will survive and thus dominate in the long run. Competition is thus linked to the concept of optimisation: via competition agents will ‘improve’ at required tasks. This framework is intuitively appealing, so much so that it can be easily exported to man-made problems and in particular to engineering and numerical optimisation. Either explicitly, as in the case of Evolutionary Computation and Particle Swarm Optimisation or implicitly, as in standard gradient-based techniques, many methods employed for computer-driven optimisation rely on some sort of competition between components of the optimisation algorithms. However, competition can express itself at different levels: a short-term winning strategy may fail in the long-term and a locally winning strategy may fail globally. Optimisation practitioners are well aware of these problems, which manifest themselves in the challenge in finding global solutions among local ones. When the process to optimise can itself respond to the optimisation and adapt to it, the dynamics can be even more complex, in which case modelling may be the only avenue for us to unravel, predict or control the process under analysis. In recent years we studied one such system. We modelled a fishery including fishing vessels harvesting several fishing zones of constant resources [3]. Vessels compete by aiming at under-exploited areas, whereby avoiding sharing the limited resource with the majority of other vessels. This is an example of a Minority Game [3] and [4] and in the Game Theory literature it is known that this apparently simple process can generate complex dynamics. Next, we included a resource dynamics by simulating population growth in the target species [5] and [6]; this imposed on the vessels the additional complication of accounting for the evolution of the resource abundance in response to fishing. Finally, we included fleet managers and resource managers, who can impose constraints of the fishery either by regulation [6] or by centralised fleet control [7]. Each step increased considerably the overall complexity of the problem. For each scenario, we studied how different fishing strategies perform both in a single strategy setting (in which all vessels in the fleet adopt the same strategy) and in an evolutionary economic setting (in which strategies spread in the population according to their past performance). Of particular interest is the comparison between the following strategies: a purely competitive approach, which we call MG in this paper (since this is the strategy normally adopted in Minority Game studies), in which agents aim to optimise their individual return, and a more collaborative approach, called Collective Intelligence (Coin in the rest of the paper), in which agents try to optimise their impact on the fleet return. It seems reasonable to consider MG as the ‘null hypothesis’ against which we test the Coin performance. One of the most important results of our previous work is that the effectiveness of Coin depends on the ratio between available resource and fishing effort. In particular, the transition from Coin to MG dominance in the fishing fleet coincides with the transition from under-exploited to over-exploited resource status (in this paper the level of exploitation is operatively defined as the ratio between available resource and harvesting potential. Other definitions, accounting explicitly for resource dynamics and other ecological factors, could also be employed, but would make the interpretation of the modelling results less straightforward.). In [5] we have speculated that this transition could, in principle, be used by a resource manager to detect the level of resource exploitation and decide on possible intervention. The purpose of this work is to explore this idea further and discuss some steps towards a possible implementation. We start by setting the problem and describing the Coin approach. We then summarise the results from our previous work most relevant to this paper and analyse our new results. We conclude with a discussion of the possible future development of this research.

We presented a method which allows agents to choose dynamically between purely competitive and collaborative strategies in a mixed-strategy approach. In a minority game-like problem, this approach shows improved performance for both centralised and decentralised decision making. The balance between agents choosing the competitive and collaborative strategy also gives an indication of the resource status and may be used for both monitoring and management. The core of the approach lies in choosing dynamically what strategy to employ to carry out a complex task in a competitive environment. Unlike previous work in [6], the choice is not carried out solely according to standard evolutionary criteria, that is by evaluating a strategy according to past performances, but also by including an information theoretical criterion, which is used to evaluate and compare the information contents of possible strategies. As a result, two dynamics are at play, each with its own time scale; at a slower time scale, those strategies will survive that have superior aggregate performances, as per traditional evolutionary dynamics. At a shorter time scale, the MaxInfo approach provides decision-makers with the strategy that gives the clearest (even if simplest) instruction on what to do. As we may expect, the two dynamics are strongly interlinked; because of the nature of the problem we address, any decision taken by an agent will affect the future behaviour of the community and thus of future resource status as well the future behaviour of the agent itself. For example, in the case of a slightly over-abundant resource, if most agents follow a Coin strategy the catch improves; thus, by following Coin, the difference between impact and contribution decreases and consequently so does the difference in information content between catch and impact. We encountered this very result in the discussion of Fig. 2b, in which the difference between information content of Coin and MG for a fully decentralised approach was unexpectedly low; it is reasonable to expect that such an information gap would have been higher, had less agents chosen a Coin strategy. The conclusion we can draw from this observation is that, while it is straightforward to see how MaxInfo affects the choice of the fishing strategy, in a less obvious fashion the choice of the fishing strategy (and thus the history of the simulation) affects the performance of MaxInfo. It is the non-trivial interplay between these two dynamics which is at the core of the performance of the algorithm we propose and more experiments may be needed to further explore its implications. In the view of a potential adoption by human agents, the information based approach seems to have some advantages against the one based on performance. The agents need to carry out more book-keeping, since records of both catches and impacts need to be stored and calculated, however the processing of this records requires few simple operations for which in principle no computation devise is strictly necessary. More important, the agents do not need to commit in advance to a specific strategy in order to evaluate its returns, nor do they need to trust reports from colleague/competitors in order to judge a strategy they have not employed; all they need to do is to check which of their own record (Coin or MG) provides a more informative instruction. This may make it easier to convince agents in the real world to adopt this approach. The second thread of the paper is the potential use of the MaxInfo approach to monitor the resource status. This also shows promising results, but leads to the question of what to do with such information should it suggest that the resource is badly over-exploited. Decision making in this scenario requires a proper model of exploitation cost, as well as alternative employment options for fishers and managers should a reduced fishing effort be necessary. We plan to analyse this in future work by casting it into an evolution dynamics scenario as in [18]. In this paper we modelled a fully renewable resource, and perfectly rational agents, which not only do not make mistakes, but also never cheat. Both assumptions are unrealistic, and have been chosen because of a ‘reductionist’ intent of separating the factors affecting a problem and evaluating the MaxInfo approach in an easy to understand test case. Clearly the approach will need to be assessed under more realistic scenarios as in [5] and [6], which we also endeavour to do in the near future. Finally, it is crucial to assess the receptivity of the approach for real agents. It is equally important to evaluate what modifications real agents may impose on the method, should they adopt it. We initially tested this in a role playing experiment with human actors and witnessed the unexpected solutions human subjects may provide to given problems. These are very hard to model and forecast in detail. More extensive experiments in this setting have already been planned and we believe are crucial to properly evaluate this approach and the best adoption path.