The model of Agent Strategy--论文代写范文精选

有四个关键问题是这个框架内探索。每个问题的担忧应该如何构造代理模型,反映了实验对象的行为。对于每个问题,我列出了一些可能的答案和解释模型来适应实现这些。下面的essay代写范文进行论述。

General simulation structure 
The structure of the computational model follows that of the Sonnemans experiment as closely as possible. Thus there are 36 experiments in parallel, each with one agent. The agents have to decide when to stop a sequence of offers and accept the value of the highest offer so far (minus the cost of the offers and a fixed ‘fee’ of 50 cents). Thus in each run of this simulation each of 36 agents plays 69 games. Thus each agent has to decide for each offer in each game whether to ask for another offer or accept the highest offer so far and at the end of part 3 to specify a fixed strategy (of the form discussed previously) for part 4. The material it has to work on is its experience in terms of its earnings and costs in previous games. How the agents specify strategy and make decisions based on this experience is the nub of the model.

Agent model Following the hints in Section 4, each agent has a small population of strategies which represent its current hypotheses. These are in the form of those in Table 1. As a result of the experience of each game the agent evaluates its hypotheses, keeps some, changes some, combines some and forgets the rest. In any particular game it uses the strategy that it has evaluated as its best to decide when to stop the bidding process. At the start each agent is given an initial population of these strategies generated at random, and from then on the agent works with these to produce newer and better strategies, so the population evolves and the strategies improve. There are four key questions to be explored within this framework. Each question concerns how the agent model should be structured so that it reflects (as far as possible) the behaviour of the subjects in Sonneman’s experiment. For each question I list some possible answers and explain how the model can be adapted to implement these.

Exactly how should the agents adapt their existing hypotheses in the light of their experience? There are many possible operations that a subject could use to produce the next population of hypotheses. I list the selection that I implemented. election. this is the operation of keeping the best existing hypothesis, as currently evaluated. 

The idea is that one holds on to one’s best hypothesis. propagation. Select a hypothesis probabilistically according to its current evaluation and keep it. Since selection is probabilistic this means that you keep a selection of your current hypotheses biased towards those that are currently doing best, but you might sometimes keep less good hypotheses as well. new. Introduce a totally new randomly generated hypothesis. 

If one does not introduce any new hypotheses there is a danger of stagnation after a while. page 10 generalisation. Select two existing hypotheses probabilistically according to their current evaluations and join them into one new hypothesis using the boolean ‘OR’ function. Thus if the selected hypotheses were ‘H≥70’ and ‘N≥7’, the new hypothesis would be ‘H≥70 OR N≥7’. specialisation. Select two existing hypotheses probabilistically according to their current evaluations and join them into one new hypothesis using the boolean ‘AND’ function. Thus if the selected hypotheses were ‘H≥70’ and ‘N≥7’, the new hypothesis would be ‘H≥70 AND N≥7’. join. Select two existing hypotheses probabilistically according to their current evaluations, randomly choose an appropriate function and join them into one new hypothesis using this function. The ‘generalisation’ and ‘specialisation’ operations described immediately above are special cases of this. 

In what way should the agents evaluate their hypotheses?
Clearly the most obvious way in which the expressions might be evaluated is according to the net earnings that using each hypothesis would have generated on the offer sequences that are encountered. However that is not the only possibility. The agent could be also biased in favour of syntactically simpler hypotheses or be in favour of those strategies that incur fewest costs. As with the operation of variation there are quite a number of ways of combining these factors in the evaluation of hypotheses. It is possible that the subjects are attempting to maximise some function of these, but it is also possible that these factors might be applied separately, for example by eliminating all those which perform badly in any of these respects. In this model each of these three aspects (earnings, costs and syntactic complexity) are evaluated for each strategy. There are then several different ways of treating the evaluation for each aspect. I chose four: a set proportion of the worst strategies with respect to each can be discarded and the rest just kept on; the strategies would have to reach a minimum level in order to be kept on; the strategies can be given a simple score with respect to each aspect; or they can be ranked.

How far back should the agents evaluate new hypotheses against past experience? 
An optimal agent would remember the offer sequence in every single past game and use all of them to evaluate its strategies. This would be implausible in this case as the subjects were not allowed to keep notes and would have had to rely on their memory. Thus there is the question of how far back subjects recall in the process of evalutating their hypotheses. Broadly, the further back a subject recalls the more efficient and less contingent is their strategy development.

What number of hypotheses should the agents hold at any one time?
It seems clear that human beings would not merely consider one strategy for each game, starting completely afresh if this one was unsuccessful. Rather it is likely that they would recall previous strategies they had tried or thought of as a basis for further strategies. Thus, in effect, the subjects will have access to a variety of past strategies at any stage. However it is also clear that the ability of human subjects to recall (or reconstruct) past strategies is limited, in other words we forget some of our past thinking. Thus what would be a realistic number of strategies for an agent to hold within this framework is also a parameter to explore. Broadly, large populations implement a more exploratory search while small populations implement a more ‘path-dependent’, contingent search. Clearly, the total number of possibilities that could be constructed by varying the above parameters is vast. However, thanks to Sonnemans’ data, we are able to pin it down in several different ways. What I will do is to exhibit the first steps in this modelling process. This is a process that obviously will need to be continued.

The results 
Many different simulations were run. Each run was set up with a different combination of operators, model parameters and evaluation methods. There is not room to display all of these here but I will summarise the results and show some of the more significant simulations in more detail. For ease of reference the two ‘signature’ graphs of the target behaviour are shown in Figure 2 below. The first is a rescaled version of Figure 1 and the second is a barchart showing the average proportion of times that subjects earlier than the optimal, at the optimal point and after the optimal point for parts 1, 2 and 3. Please note that whereas Sonnemans’ calculated the expected average earnings and spread of earnings using 160,000 simulations of randomly generated games, I had only the computational resources to do relatively few (400 simulations). Thus while the diagram in [2] are accurate to one decimal place (with a probability of 95%), mine are only accurate to the nearest 1 (in the mean earnings), and will overestimate the expected spread of the earnings.

Discussion 
Clearly the model exhibited represents only a step towards an accurate description of the target behaviour observed by Sonnemans. The guidance from cognitive science as to the nature of the learning processes involved is, at present, only suggestive. This results in there being considerable uncertainty about the nature of the learning process, uncertainty that the current observational data does not completely resolve. What the model does do is to point up what information we are lacking about the target behaviour that would enable us to improve the model. In this case one thing that is missing is any direct information about how the subjects are developing their models as they learn.

If these were done the process of descriptive modelling could go through another, more accurate iteration. This might highlight more facets that would require experimental or observational investigation. Alternatively it might reveal that some of the inevitable assumptions we had made about the learning process were unwarranted. But in either case we would have learnt something about the actual learning processes involved.

51Due网站原创范文除特殊说明外一切图文著作权归51Due所有；未经51Due官方授权谢绝任何用途转载或刊发于媒体。如发生侵犯著作权现象，51Due保留一切法律追诉权。
更多essay代写范文欢迎访问我们主页 www.51due.com 当然有essay代写需求可以和我们24小时在线客服 QQ:800020041 联系交流。-X