This week I have focused my attention on the max-sum algorithm. I have implemented it in java and have created a simple visualisation to show it trying to perform graph colouring. I am still having problems with the implementation as it doesn't seem to converge to the correct solution. It will often reach the correct solution, but if left running will then converge to all nodes coloured one particular colour. I make sure each variable to function message is normalised before it is sent out and I also randomly allocate a variable state preference to each agent before the algorithm is run. In cyclic graphs (which will always be present when using agents) the max-sum algorithm does not guarantee convergence to an optimal solution, or a solution at all therefore I wonder whether I should implement the termination constraints and see if this makes any difference.
Today I have been reading a number of papers:
"The contract net protocol: High-level communication and control in a distributed problem solver" - presents an algorithm for task execution in a decentralised way using agents. A Manager agent will request a task to be completed, agents submit offers to complete that task with various constraints, the manger agent chooses the best offer, the offers owner becomes a contract agent and completes the task. Note this is a simple algorithm and there is no counter offers allowed.
"Intelligent distributed autonomous power systems (IDAPS)" - Presents IDAPS "Intelligent distributed autonomous power system" which is a form of microgrid.
I also attended the first Energy meeting where we discussed the particular areas we are all working in. These meeting will allow us to go into detail about particular problems that we have, or in order to get people interested in collaborating in a particular area. These meeting differ from the ORCHID meetings in that they are specifically in the area of energy and not the wider area of agents.
No comments:
Post a Comment