Definition of Strategies

In order to achieve flexible agent behaviour, we have to conceive about the way the agents would be designing their own strategies. Some part of the literature (D2.1) refer to strategy as being a mapping from the current state to a current action for the agent. If this is an ideal manner to design the implementation, then the modellers have to provide a list of actions or functions that are applicable in one state of an agent. There would also be a separate selection procedure which would allow these agents to choose between these actions.

Another manner might be taking the hierarchical approach, where a defined task may be divided into smaller units or targets which have to be done to reach the ultimate goal. The agent could then choose between these small tasks to achieve its goal.

The question currently on my mind is, if we were to follow the above approach would a strategy be confined to every particular state or should a strategy be equivalent to the function map, which would include all the roles the agent has and all its functions and how these would be arranged accordingly to achieve maximum profit for itself for every iteration.

This would allow the agents to review their day to day activities, and possibly evolve their behaviour from the old practices into easier ways of performing activities. Also show what the agents would do if some exogenous shocks are introduced into the system.

The key we would target for: modellers should basically give the basic functions the agents can do, the agents would then use this information and looking at what is advantageous to them change or optimize their behaviour. Remember to consider the problem of time activation for some of the agents.

Contradiction to mechanism design in economics

Mechanism design is one facet of game theory which creeps up in multi-agent systems literature when complex decisions have to be taken. The 2007 Economics Nobel prize was awarded on the same theory being proposed to introduce efficent market algortihms.
In theory, mechanism design relates to the making of decision by every agent keeping in consideration what the other agents might decide, not only to maximise its own utility factor but also the system's overall performance. And this is why most MD literature refers to auction based mechanisms where every agent submits their bids, and a central agent decides who gets to perform the task. There is much research progressing towards how this could be applied on a decentralised manner to acheive the same goals.
However, in reality if we were to apply this principle to for say, malls competing for the same pool of customers, every mall agent would be wanting to increase their own profits, possibly predicting the other mall's behaviour before making the decision, giving less consideration what would be better for the complete society. This is possible why the proverb, rich get richer and poor get poorer might be true. every agent behaves only on their preferences, not caring how the community gets affected. And this is why some businesses might be able to monopolise the business by various strategies taken at the right time while others only struggle to survive.

Deviation- back to FLAME methods

Due to Eurace visit, have to stop work on state generation and turn back to FLAME methods to code up the ipd game example. Probably later I can use this coded example and modify it to states in a month's time:).

The drive towards event-based agents

Flame allows the iterations to be represented by the shortest time scale a model could possess. This creates problems for issues where agents would be reactive and accordingly change their behvaiour during the course of the model. A way to inhabit this aspect is a turn towards state based modelling of agents. Every iteration would then be represented as the path from the start state of the agent to the end state of the agent in that one iteration. This would map behaviours of agents with deliberative and reactive natures corrently into the model and can also see a possible extension to the hierarchy of various models embedded into each other. Consequently, this raises the problem of parallelisation of agents. This leads me to question whether we need to have synchronisation points or if we were to remove them would deadlocks still arise? If so how would this be handled. Lack of correct use of formal definition of agents in Flame is causing discrepancies!