WHAT ARE PLANS FOR Philip E Agre David

WHAT ARE PLANS FOR? Philip E. Agre David Chapman October 1989 CS 790 X ROBOTICS Presentation by Tamer Uz

WHAT IS A PLAN? • A road map to the goal? • Is it a well organized set of instructions? • Can just a single if statement considered to be a plan?

TYPE OF PLANS UNDER THE MAGNIFYING GLASS OF THIS STUDY Plans As Program View Plans As Communication View

PLANS AS PROGRAMS • Plan use is simple a matter of execution of the instructions • Executing the plan is just simply walking over it in a syntactic and mechanical fashion • Directly determines the user’s action • No interpolations, substitutions or rearrangements

ISSUES with the PLANS AS PROGRAMS • Complexity in both representing the world and plan making • If the planner fails to anticipate an uncertainty, then there is no way to recover, everything will fail. • Some hierarchical sub plan methodologies are offered to tackle with the uncertainties. But it is not plausible to make a hierarchical structuring on something that is unknown. • The executive has to establish a connection between the concrete situation and what the plan says for it. So to be able to make this connection the executive has to get the label the items around so that the communication with the plan makes sense. But often there won’t be labels on the items.

A NEW PHILOSOHY SHOULD BE FOUND Which can help overcome the unpredictable brutalities of the wild world Which does helps the planner concentrate on the way of achieving the goal, rather than taking into account all those unpredictable outside conditions. And it should be easy as in “Boil the water, prepare the coffee put it on the table”

SOME MORE THOUGHT • Is the agent supposed to execute the instructions in the plan like a soldier, or would that be OK, if it uses the plan just like it’s other resources and make use of it? • The authors claim that the plan should not be something central as in the case of plan as programs. I don’t know how this would be, because even though the plan is rough set of initial instructions, it is still the route to the goal even if it changes along the road, isn’t it? • OR SHOULD NOT THERE BE ANY PLAN AT ALL? How about just situated instruction use to achieve the goal?

PENGI EXAMPLE A simple run and chase game. • I am not the author’s advocate. • Looks like an irrelevant example. • They say that a traditional plan as a program view would have done dumb things, without even using simple if decision branches after each clock step. • And they say that their little funny Pengi does not use any plan? I could not understand this? Apparently it does not move aimlessly, so there should be some kind of a plan in it’s operation. • Did anybody understand, how it works “without a plan”?

PLANS AS COMMUNICATIONS • Like getting plans from somebody else, or making plans of our own. • Without thinking of any unexpected situation. • Easily modifyable, or sometimes can be totally abandoned and replaced with some other plan?

ASSUMPTIONS for PLANS AS COMMUNICATIONS • Indexicality: The interpreter of the plan knows what the planner is talking about. Planner will not pinpoint the objects in the scenario unless it is particularly required. In other words changing meanings according to the context. • Projection: What if I don’t do this instruction? • Reflexivity: To be able to know what the other party knows.

CONCLUSION • Plan use is not necessary for sensible action • Plans represent roughly the activity they describe, rather than in the way a program represents the computation it describes. • Matching a situation with the one described in the plan is a huge pain in the neck. • The ability to make plan and use it comes from the experience of cooperative language use in the context of the activity. • Plan use relies on unbounded set of assumption, that the plan maker and the user have to share. • Using one’s own plan is much like plans communicated by someone else.

SOME OTHER ALTERNATIVES • Interleaved planning • Behavioral Modules • Heuristic Planning • Reactive Planning • Plan-as-Constraints

INTERLEAVED PLANNING • The planner makes its plans as always; when the executive gets into trouble, it passes the control back to the planner, which assesses the situation and makes a new plan. ISSUES • In reality, the executive has a little or no access to the reasoning behind the plan. That is why it is almost impossible for the executive to invalidate the plan. • It does not detect the trouble, unless it is too late. • It does not provide a way registering unexpected opportunities.

BEHAVIORAL MODULES • Interact in complex, but well understood ways with the physics of the domain to effect specified conditions. ISSUES • Issues are not covered

HEURISTIC PLANNING • It refers to the heuristics that would help search the plan space. ISSUES • It is domain dependent. • The domain should be simple and specific.

REACTIVE PLANNING • Conventional executive together with an externally generated plan library. • Choose one of the plans from the library as the circumstances evolve. ISSUES • System efficiency depends on how big the individual plans are. • The selection algorithm needs to be explained.

PLANS-AS-CONSTRAINTS • In this view, plans are set of beliefs and intentions, which are constraints on possible features • The view does not specify how these constraints effect the action, but they are used to interpret other agents’ action. ISSUES: • One of the authors will study on this and discuss the relationship between this and plan-as-programs and plan-as-communications.