This shows you the differences between two versions of the page.

Link to this comparison view

hcmi:brian_pendleton_sumpter [2014/08/13 16:16] (current)
tlund1 created
Line 1: Line 1:
 +===Control methods for self organizing systems===
 +* '''​Misc notes'''​
 +** In a large group, many behaviors are specified in terms of interactions instead of single agent goals.
 +* '''​Integrity'''​ - The goal is not to micro manage every action of the robot, but to assign high level behaviors or objectives to be executed. ​ To be successful, the robots must maintain some degree of autonomy. ​ The robot should think twice before following the users instructions to drive off the cliff :).
 +* '''​Variability'''​
 +** Steer evolution of desirable traits. ​ The agents can cover the search space of desirable behaviors faster if they do different things.
 +** Reinforcement learning.
 +** Different response thresholds. ​ Allows for a smooth behavior transition in a team setting. ​ For example, as the problem grows in size, more agents are automatically recruited (due to having varying response levels).
 +** In some tasks, variability is desired or required to accomplish it.  Having individual variability saves the operator from specifying specific behaviors to each robot. ​ They can self organize to complete a high level task.
 +* '''​Positive Feedback'''​
 +** Recruitment. ​ Can get n agents to do a task without explicitly assigning n agents. ​ Vital for scalability of the system from an operational perspective.
 +** Allows control of large groups without controlling large numbers of individuals.
 +** Allows operator to shape the behavior of the group. ​ Reinforcement.
 +** Agents automatically reinforce positive behaviors. ​ Allows operator to specify high level behaviors (such as laying trails) that agents then use.
 +** Hysteresis effects. ​ Allows to group to "latch onto" certain behaviors and resist perturbations of the system. ​ Allows groups of agents to reach a steady state, allowing the operator to go focus on something else.  Allows for memory in the system. ​ Allows large groups to lock into certain behaviors.
 +** Operators can manipulate thresholds, increasing or decreasing the likelihood that a large global behavior will be adopted.
 +** Allows information to quickly propagate through a group. ​ "​Because you're doing it, I'm going to do it to" ​ Allows information to pass between agents even though the operator hasn't (or can't) send the information directly.
 +* '''​Negative Feedback'''​
 +** Negatively reinforce undesirable behaviors to shape group behavior
 +** Decrease the effect of perturbations of the system.
 +** Allows operators to hold the group in a certain state.
 +** Reverse the effect of a previous Hysteresis change. ​ Tell the group to "look for new options"​
 +* '''​Response Thresholds'''​
 +** Stabilizes the group by requiring some consensus before a certain behavior is adopted.
 +** Operator can directly (or indirectly) modify thresholds the perceived environment of the agents increasing or decreasing the likelihood that a certain behavior will be adopted.
 +** Agent can notice that it's not behaving like the rest of the group and comply.
 +** Provides a way of specifying the importance of the task.
 +** Provides a way to to specify how many agents should respond (individual thresholds vary from agent to agent).
 +** Information propagation mechanism. ​ Once consensus is reached, information propagates quickly.
 +* '''​Leadership'''​
 +** One agent (the operator or another agent) must be able to communicate information and actions to the rest of the group. ​ Agents must not go with the flow all the time.
 +** Allows for agent autonomy. ​ Agents can try out new ideas. ​ Good ideas recruit others and pass information throughout the group. ​ This decreases the overhead for the operator.
 +** Simplifies control problem for operator. ​ Operator interacts (directly or by means listed above) with small group of leadership prone agents that then propagate information or recruit other agents.
 +** Stabilizes positive feedback. ​ Allows agents to split tasks among group and transition from one task to another. ​ Positive feedback alone would cause the whole group to quickly change behaviors. ​ Leadership within groups provides a stabilizing force that allows the group to break into subgroups.
 +** Provides group coordination. ​ Allows an agent to go "​fetch"​ others and bring them to the group.
 +* '''​Inhibition'''​
 +** Another information propagation mechanism. ​ Allows the operator to say "​don'​t do this", or "​don'​t do this as much" or "this area is off limits"​
 +** Informed agents or leaders (or operators) can steer the group away from undesirable behaviors and towards desirable ones.  "​don'​t go that way, go this way"
 +** Provides another stabilizer for positive feedback.
 +** Provides group stabilization. ​ Positive feedback inhibits deviance from the current behavior.
 +* '''​Redundancy'''​
 +** Decreases operator burden by ensuring a task will be carried out, even if the original agent responsible is unable to do so.
 +** Makes the group more resilient to disturbances. ​ Stabilizes behaviors. ​ If the leader that originally initiated the action is knocked out, the group will still continue the behavior (assuming the system is in a stable state). ​ The next most leadership prone individual may take charge.
 +** Allows an operator to pick certain agents to perform tasks. ​ Taking away one agent will not change the global behavior of the group. ​ Allows the operator to focus on one agent without worrying about the rest of the group.
 +* '''​Synchronization'''​
 +** Synchronizing large numbers of individual elements is hard to do (try to synchronize 20 pendulums by yourself :)).  Automatic synchronization of tasks frees the operator from micromanagement.
 +** Group forming - allows operator to command a group instead of individual units, simplifies operator control problem.
 +** Reduces need to closely monitor agents in critical situation. ​ For example, coordinated surveillance or attack
 +** Agents can synchronize with operator controlled agent or operator actions. ​ "do this"
 +** Operator can establish patterns and routines that the agents follow. ​
 +* '''​Selfishness'''​
 +** Provides a reward structure. ​ Operator can modify agents using rewards and punishments.
 +** evolution models
 +** Provides a framework for individual initiative. ​ Brings in utility theory.
hcmi/brian_pendleton_sumpter.txt ยท Last modified: 2014/08/13 16:16 by tlund1
Back to top
CC Attribution-Share Alike 4.0 International
chimeric.de = chi`s home Valid CSS Driven by DokuWiki do yourself a favour and use a real browser - get firefox!! Recent changes RSS feed Valid XHTML 1.0