Chapter 6: Applying the PSSH Overview • Explain why ID3 and SHAKEY both count as illustrations of the heuristic search hypothesis • Introduce frame problem • Introduce background to the language of thought hypothesis (as an application of the PSSH) Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 ID3 and PSSH • ID3 transforms one highly complex symbol structure (database) into another (a decision tree) • The decision tree itself works by transforming symbol structures according to rules • Rules are the IF. . . THEN . . . rules built into the decision tree Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 Heuristic search • Problems are solved by generating and modifying symbol structures until a solution structure is reached • Problem-spaces are generally too large to be searched exhaustively (brute force algorithms) • Search must be selective heuristic search rules Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 ID3 as heuristic search • ID3 is a tool for navigating through the search space of decision trees • The algorithm that it uses sorts through possible decision trees using measures of information gain and entropy reduction Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 Basic concepts Entropy – level of uncertainty (relative to a target attribute) Information gain – reduction of entropy (relative to the target attribute) The algorithm compares different classifying attributes in order to determine which has the highest information gain Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 SHAKEY and logic programming Basic ideas (1) states in state-spaces characterized by sets of predicate calculus formulas (2) problems are conjectures to be proved (3) problem-solving is finding proofs using prooftheoretic techniques and heuristic search techniques Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 Turn on the light switch GO SHAKEY! 1. Goal wff = STATUS (SWITCH1, ON) 2. STRIPS solution = {goto2 (BOX1), climbonbox (BOX 1), Climboffbox (BOX 1) Pushto (BOX 1, SWITCH 1) Climbonbox (BOX 1) Turnonlight (SWITCH 1) Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 The frame problem • Applying an operator to a state in a state-space problem will affect some relations between objects in the state but not others • How is the system to keep track of which relations have changed and which not? • Proofs would be very complicated if there were axioms for each operator specifying the relevant changes and nonchanges Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 Possible framing axiom: (on C B) & (put B table) (on C B) Framing axiom? Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 SHAKEY and the frame problem SHAKEY has a limited solution to the frame problem built into it The F-rules in STRIPS have three components • Precondition formula • Delete list • Add formula The Delete list makes explicit how the state description needs to be altered Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 Example • Pickup (x) • Precondition formula: ONTABLE (x) & HANDEMPTY & CLEAR (x) • Delete list: {ONTABLE (x), HANDEMPTY, CLEAR (x)} • Add formula: HOLDING (x) Cognitive Science José Luis Bermúdez / Cambridge University Press 2010 Limitations of SHAKEY • SHAKEY’s explicit solution to the frame problem works only because the microworld is so restricted • Each operator will only have a small number of effects on the state • These can be explicitly listed • But this won’t scale up, when we introduce other agents and a more complex environment Cognitive Science José Luis Bermúdez / Cambridge University Press 2010