There are fundamental costs to starting a task and continuing a task. Setting up the tools you need to begin working has a fixed cost that does not depend on how long you work. So there are good reasons for splitting up your time into long segments spent on particular sorts of work, rather than switching modes rapidly.
In multi-agent systems—or, as a special case, in dynamically inconsistent “agents”—this effect will be (vastly?) magnified. Say Bob is in a group, and he wants the group to do task B, but the group is leaning towards doing task A. Once the group has started doing A, it will be much harder to convince them to switch, because they will have already paid the starting costs of A. So if Bob wants to make a case for doing B, the best time to do it will be before the group starts doing A. And since Bob knows this, he has an incentive to filibuster for as long as it looks like he will lose the vote.
Thus, multi-agent systems will have a tendency to spend more time doing nothing than dynamically consistent agents.
Here’s a theory I have.
There are fundamental costs to starting a task and continuing a task. Setting up the tools you need to begin working has a fixed cost that does not depend on how long you work. So there are good reasons for splitting up your time into long segments spent on particular sorts of work, rather than switching modes rapidly.
In multi-agent systems—or, as a special case, in dynamically inconsistent “agents”—this effect will be (vastly?) magnified. Say Bob is in a group, and he wants the group to do task B, but the group is leaning towards doing task A. Once the group has started doing A, it will be much harder to convince them to switch, because they will have already paid the starting costs of A. So if Bob wants to make a case for doing B, the best time to do it will be before the group starts doing A. And since Bob knows this, he has an incentive to filibuster for as long as it looks like he will lose the vote.
Thus, multi-agent systems will have a tendency to spend more time doing nothing than dynamically consistent agents.