|
||||
|
Adaptive negotiation strategy for agents in electronic markets
| Permalink A common requirement of distributed multi-agent systems is for the agents to be able to negotiate an agreement or settlement among themselves, for example in applications such as load balancing, resource allocation or job scheduling. A typical way to reach such agreements is through auction mechanisms (see Figure 1) such as the continuous double auction (whereby a number of buyers and sellers simultaneously offer to buy and sell the same good), which is employed in many modern financial markets. There can also be semi-competitive trading environments that pit agents against each other in certain respects, giving them an incentive to act self-interestedly, but where agents can also sometimes increase their utility through cooperation. Yet another example of a settlement strategy is a variation on the El Farol bar problem (which asks how each individual can autonomously determine whether or not to go to the bar, with the selfish goal of maximizing his or her own payoff, assuming that everyone would rather stay at home if the bar is too crowded but would prefer going to the bar otherwise) called the minority game,1 wherein an odd number of players independently choose one of two options, and the set of players who choose the least popular option are the winners. Figure 1. Auctions are a common mechanism for arriving at settlements in multi-agent systems. 
All of these negotiation scenarios have several features in common. First, they require agents to deal with different dimensions of uncertainty, such as uncertainty about the future. In a continuous double auction, the buyers have to decide whether to buy now at this price, or wait for a lower price with the risk that the current option will be taken by someone else, forcing them to buy later at a higher price. Second, the agents must be able to dynamically adapt their attitude to risks. For example, a buyer who has not succeeded in buying the good for many rounds should become more risk averse and so more willing to offer higher bids. Third, the agents should be able to discard superseded facts since, in a dynamic setting, not all knowledge is equally useful. When circumstances change, some information may become entirely useless or even harmful if used, because it is no longer relevant. A buyer in a continuous double auction might fail to buy any good whatsoever once the market has become competitive, if he still lives in the ‘good old days’ when everything was always sold at very low prices. These requirements of a negotiation environment are addressed by rational agents that make decisions with the goal of maximizing their utility functions. However, agents called upon to enter a ‘decision-requiring environment’ whose conditions are not completely clear may not have enough time to learn an optimal strategy for the utility-maximizing function (unless losing money is a goal). Even if they do have time to learn a strategy, the dynamics of the market are such that conditions may rapidly change and render it redundant. Negotiation scenarios are also characterized by the nature of interactions among agents. In a semi-competitive trading environment, agents may employ deception or ‘lying’ to maximize their outcomes. Players in such games have to be aware of the existence of strategies involving deception, and they should likewise know that signalling systems can be used to misdirect, and that there is both a cognitive and an economic dimension to trust decisions in such contexts.2 For example, Player A might ‘trust’ Player B more than Player C, but if the payoff for taking the course of action advised by Player C is very much greater than that for following Player B, then A might—depending on its attitude to risk—choose the less trustworthy option. Of course, agents must also be aware of the consequences for their reputation and trustworthiness in other players' eyes before they decide whether to tell a lie. The agents also need to be cognizant of the limits of their knowledge, and recognize when information previously acquired is no longer usefully contributing to the decisions that they have to make. In other words, agents must be able to ‘forget’ information that has outlived its usefulness. In this connection, some neuroscientists have claimed that the capacity to forget is crucial for the efficient functioning of the human mind,3 and it has been argued that forgetfulness is an important factor in human social, cultural and psychological development.4 Finally, agents should be aware that many of the decisions they have to make are actually approximate and fuzzy, rather than rigid and precise. For example, an agent that decides to buy an auctioned item for $100 would in most cases have been (almost) equally happy to buy it at an asking price of $101. We can therefore say that, in practice, the agent decided to buy the item for ‘around’ $100, rather than for precisely that amount. Multi-agent research has demonstrated the technical feasibility of achieving these types of adaptations through dynamic situation awareness. For example, in our previous work we represented agents' risk attitudes as a real number between 0 and 1, which was adapted dynamically5–7 by incrementing and decrementing its value in response to trading outcomes. The same techniques have also been used in other situations, including combinatorial auctions8 and resource allocation.9 Fuzzy logic is applied at different stages of the decision making process, and the fuzzy decision is defuzzified (translated into a quantifiable value) before it is applied to the environment. Algorithmic trading is widely used in financial markets and other spot markets (such as those for oil and electricity), where it is an essential tool for minimizing risk and providing liquidity. Elsewhere, cloud computing is increasingly evolving toward differentiated service offerings, such as software-as-a-service and platform-as-a-service. All these fields of application will increasingly call for processes—i.e. agents—capable of trading and negotiating in electronic markets. Our model for adaptation through dynamic situation awareness will have an impact on agents' ability to effectively participate in such markets, because it offers a generic strategy for any situation rather than an optimal strategy tailored to specific circumstances. Furthermore, if all the agents are adaptive in the manner described, the market itself becomes more efficient and more robust, which translates into an ‘all-win’ situation. In future, we envisage continuing this work by investigating how the principles employed in multi-agent systems can be applied in real-life trading situations, possibly after some refinements, and by developing a formal model (i.e. a calculus) that enables us to better analyse those principles. References
Stay Informed
|
||||
|
||||