Learning and exploiting knowledge in multi-agent task allocation problems

Imagine a group of cooperating agents attempting to allocate tasks amongst themselves without knowledge of their own capabilities. Over time, they develop a belief of their own skill levels through failed attempts at completing the tasks they are assigned. How will various task allocation approaches perform when there exists this added level of complexity? In particular, we compare two task allocation strategies: a greedy, first-come-first-serve approach, and a more intelligent, best-fit method. By varying the number of tasks along with the amount of time it takes to complete those tasks, we find that the different task allocation methods work better in different situations. Because of the way the tasks are allocated by the two methods, the greedy approach does a better job of giving agents opportunities to learn their capabilities. Thus, the greedy approach allows for quicker learning and performs better on problems where the task durations are short, whereas the best-fit method performs better on problems where the task quantity and durations are large. What is needed is a hybrid method that balances between the exploration of the greedy approach and the exploitation of the best-fit method.