Learning Plans
- From the existing rules of STRIPS we can use Explanation Based Generalization to obtain new rules.
- The new rules which are learned can help in designing or achieving more complex plans.
- If by generalization of the rule there is a reduction in planning effort or cost then it can be saved.
- We can generalize the STRIPS rules. It can be explained with the help of an example.
- Example: The initial state is, P is on Q is on R the goal is to place P, Q on floor.
- The STRIPS rules like move(P, Q, F) and move(Q, R, F) can be used to do this.
- But saving this plan as new STRIPS rule may not be worthy. Because generalizing it may help to move any of the two blocks on to the floor regardless of their names.
- In order to generalize, the plan variables like x, y, z can be used instead of P, Q, R but the target floor cannot be generalized.
- The information about the preconditions and effects of the plan are used in plan generalization.
- A triangle table can be used to represent this information. For example, we considered the triangle looks as shown in Fig. 3.10.18.
Fig. 3.10.18 A Triangle Table for disc Unstacking
- The columns of the table has operator as their headings. Start and finish are used to represent order of plan.
- The cells that are present below the operator represent what it has added.
- The cells to the left of operator represents the operators preconditions.
- The labelling of the cells is done as, its row is represented as i and column position as j i.e., as (i, j).
- The cells to the left of finish operator contains literals which satisfy goal condition.
- Now we replace all the constants like P, Q, R with variables.
- The preconditions can be checked by a process by substituting and if they are met then the operations are performed.
- The generalized plan is known as MACROP which helps in creating longer plans.
- The literals which are present in the column1 of the MACROP are the pre-conditions and literals in row three are conjuncts to be added.
- The table that is generalized is shown in Fig. 3.10.19.
Fig. 3.10.19 A generalized Plan Schema for Disc Unstacking
- It might also be useful to construct specific plans in order to reduce future planning by providing control-strategy information like in achieving On(P, F) ∧ On(Q, F), On(P, F) has to be achieved first is the control-strategy information.
- Knowing the effects of rules is also a type of learning because, if the effect of rules is not known we can't apply them in the correct situations.
