Russell and Norvig, Chapter 27:  AI: Present and Future

27.1 Have We Succeeded Yet?
- intelligent agent design
  - percepts and actions available to agent
  - goals that agent's behavior should satisfy
  - nature of the environment
- do we have the tools to build a general-purpose intelligent agent?
  - knowledge-based deliberation vs. reflex
    - compilation generalizes deliberation into reflex
  - uncertainty reasoning
    - need to combine logic and uncertainty
    - need to combine planning and uncertainty
  - real-time AI
    - anytime algorithms
      - solution improves gradually over time
      - but always available
      - e.g., iterative deepening
    - decision-theoretic meta-reasoning
      - decide what computation to perform
        - based on cost versus benefit tradeoff
  - learning
    - success and understanding decreases with representation complexity

27.2 What Exactly Are We Trying To Do?
- rational agent
- perfect rationality
  - classical decision theory
  - maximize expected utility
  - computationally infeasible for complex environments
- calculative rationality
  - will eventually produce a rational choice
    - as defined by some calculation
    - may take a long time
- bounded optimality
  - behaves as well as possible given computational resources
  - expected utility no worse than any other agent under same constraints
  - most feasible choice for intelligent agent rationality

27.3 What If We Do Succeed?
- legal responsibility
- terminator
- AI is good, people are bad