Next: Planning and Acting with CM
Up: Multiagent Learning through Collective Memory
Previous: Collective Memory Task Environment
Consider the problem of CM in the context of MOVERS-WORLD. In MOVERS-WORLD, the task is to move
furniture and/or boxes from a house into a truck (or vice-versa). MOVERS-WORLD has several agents
who are differentiated by their types: some are ``lifters'' and some are ``hand-truck operators''.
Agents with the same type still have differing capabilities due to attributes such as strength.
There are type-specific operators that the agents can use and there are general purpose operators
that any agent can perform. The agents in MOVERS-WORLD do
not engage in any communication at planning time. Rather, they plan independently, act
independently, and only communicate to establish some cooperation when necessary.
CM can impact the performance of agents in MOVERS-WORLD in two ways during step 3a above:
- An agent can use her past successful problem solving experience to guide her in similar
situations rather than to have to plan from scratch to find a solution. An example of a
cooperative routine useful for future reuse is for a lifter to learn to load boxes onto available
hand-trucks before leaving a room.
- When an agent has to plan from scratch, her experience in interacting with the domain will
allow her to construct a plan that is more likely to succeed, since she has a better idea of both
her and other agents' capabilities. An example of improving planner performance based on
experience is that if agent A requests help from agent B and receives a negative response, then
A's planner will be less likely to construct plans requiring B's assistance in similar
circumstances.
Figure 1: COBWEB tree for calculating operator probabilities.
Andrew Garland
Thu Apr 9 11:37:41 EDT 1998