90 likes | 107 Views
Explore the concepts of memory versus computation, state-space graphs, searching explicit state spaces, and feature-based state spaces in the design of agents that plan, based on Nils Nilsson's book. Understand the role of reactive agents, computational processes, and the prediction of actions and consequences in different world models.
E N D
K. V. S. Prasad Notes for TIN171/DIT410 (Friday, 26 March 2010) Based on Nils Nilsson, “Artificial Intelligence: a new synthesis”, Morgan Kaufmann Publishers, 1998 Agents that plan
Agents that plan • Memory vs computation • State-space graphs • Searching explicit state spaces • Feature-based state spaces
Memory vs computation • Reactive agents • Do very little computation • actions selected by designer, or by learning, or genes • implemented in tables, or rules, or circuits • need lots of memory • Designer must anticipate all situation/reaction • Can the agent compute what the designer would? • Agent would then need more time but less space
Computations by the agent • Designer must specify these • not carry them out • E.g., predict the consequences of possible actions • If these can be learnt, or evolved • Agent does things designer did not anticipate • To predict consequences, agent must model • World • actions
World states and actions • Blocks world, make (ABC) from (A)(B)(C) • World modelled iconically • Actions by before-after pairs • move(A,B) • takes (A)(B)(C) to (AB)(C) • Takes (A)(BC) to (ABC) • Doesn't apply to (BA)(C) • So we can look ahead one step. To goal?
State-space graphs • Nodes = world states • Arcs = actions • (A)(B)(C) –move(A,B)--> (AB)(C) • In blocks world, all actions reversible • (AB)(C) –move(A,T)-->(A)(B)(C) • Where T is the table • So agent can see: • to go from (A)(B)(C) to (ABC) • do move(B,C), move(A,B)
Plan = path in state-space graph • Any of the nodes can be the goal • Sequence of actions needed (the plan) • Becomes path from initial to goal • Assumptions • Can represent all relevant world states and actions • No uncertainty in effect of actions • No other agent to change state • Then no sensor needed while acting
Searching explicit state-spaces • Start node is labelled 0 • Propagate numbers in waves along arcs • So arc labelled 3 has 3 step path back to initial • Wave go breadth-first; other sequences possible • Continue till you hit the goal • For single goal, can also back from goal to init
Feature-based state spaces • Nodes now labelled by features • Feature = logical proposition • Then what are actions? • STRIPS says an action is a triple • Precondition (what must be true for action to be possible) • Add list (what features become true after action) • Delete list (what features become false after action) • We are no longer in iconic world states!