Outline
• Time, Schedules, and Resources
• Hierarchical Task Network Planning
• Planning and Acting in Nondeterministic
Domains
• Multi Agent Planning
Time, Schedules, & Resources
• Basic Planning:
– What to do and in which order?
• Real World:
– What an When to do? + Limited Resources.
– JOB SHOP SCHEDULING
Limited Resources
• Resources:
– Consumable vs. Reusable.
• Notation:
– Aggregation
– Immediate Effect
– Resource:R(k)
• Requirement / Temporary Effect
Limited Resources
• No General Approach (NP-Hard)
• Just Order the task so that the
requirements are met.
• Heuristic:
– Minimum Slack Algorithm:
• Give more priority to the task with least
remaining slack.
Hierarchical Planning
• Hierarchical Task Network:
– At each “level,” only a small number of
individual planning actions, then descend to
lower levels to “solve these” for real.
– At higher levels, the planner ignores “internal
effects” of decompositions. But these
have
to be resolved at some level…
HTN Sample (cont)
HTN Cons and Pros
• What’s Bad?
– Recursion?
– Sub Task
Sharing:
• Enjoy
honey
moon in
Hawaii and
raise a
family.
• Library:
– Enjoy
Honey
HTN Cons and Pros
• What’s Good:
– Almost all real applications are HTN + some
thing else.
– It’s a heuristic to decrease the branching
factor by a great level.
NonDeterministic Domains
• What if we don’t know all about situations
and effects.
• E.g.
– Init: A table and a chair of unknown colors.
– Goal: A table and a chair of the same colors.
– Condition: Painting may have flaws.
Sensorless Planning
• We don’t know all beforehand and we
can’t find it out, even when it is done.
– Plan so that to reach the goal state,
regardless of everything. (Coercion)
– Not always possible.
Conditional Planning
• We can check the state ahead, then
perform the pre-planned program.
– Sense Actions
– Conditional Branches
Conditional Planning in
Fully Observable
• Vacuum World:
Domains
– Left: AtRight AtLeft AtRight
– Left: AtRight
(AtLeft AtRight) (AtLeft AtRight)
– Suck: when AtLeftCleanLeft
when AtRightCleanRight
– Left: when AtLeft CleanLeft
when AtRightAtLeft AtRight
State Space:
Partially Observable
Domains
Partially Observable
•
Domains
Easiest Approach:
– Assume set of current states and the next state
sets are created, quite similar to non-
deterministic actions case.
Execution Monitoring and
Replanning
• Check if the plan is going on is pre-
decided? If not, replan based on current
situation.
Continues Planner
• Keep planning, sensing and executing…
– Which is not unlikely, such as maintenance
planning, auto-pilot, plant control, …
Continuous Planner
• POP + …
– Missing Goal:
• A new goal has erupted. Just add it.
– Open precondition:
• An action has lost its support links. Add a new
causal link.
– Causal Conflicts:
• A causal link is suddenly threatened. Choose an
appropriate ordering.
Continuous Planner
• POP + …
– Unsupported Link:
• A link from start to something has suddenly last
its true value. Remove it.
– Redundant Action:
• An action no more produces something needed.
Remove it.
Multi Agent Planning
• When there is more than one agent in the
scene.
– Competitive
– Cooperative
• Coordination
– Communication
Cooperation
• Multi Body Planning
– One is in charge of all decisions…
• Having the agent as one of parameters:
– Go(R2D3, Right) ^ Go(C3PO,Left).
• Synchronization and Timing…
Cooperation – Multi Body
• Joint Planning:
– Planning using action pairs:
• Exponentially Many Actions: Actions Agents
– Having Concurrent Actions List
• Which actions happen together and which not,
such as orders in POP.
Cooperation - Coordination
• Accepting a prior Convention.
– Everyone drive on his/her right side of the
road.
– Domain Independent:
• Choosing the first feasible action.
• Producing all possible feasible actions and choosing
the one which stands first in alphabetic order!
Cooperation – Emergence
• Evolutionary Emergent Behavior
– Birds Flocking:
• Separation
• Cohesion
• Alignment
– Ants.
Coop. - Communication
• A short message expressing
– the plan / next step.
• A message expressing the next step.
• Plan Recognition!
Competition
• Minimax + Conditional Planning
