Presentation is loading. Please wait.

Presentation is loading. Please wait.

CSE5900 Lecture 9AI & MM Pt. 11 AI and MM Part 1 (An Embarrassingly Over-Simplified Introduction)

Similar presentations


Presentation on theme: "CSE5900 Lecture 9AI & MM Pt. 11 AI and MM Part 1 (An Embarrassingly Over-Simplified Introduction)"— Presentation transcript:

1 CSE5900 Lecture 9AI & MM Pt. 11 AI and MM Part 1 (An Embarrassingly Over-Simplified Introduction)

2 CSE5900 Lecture 9AI & MM Pt. 12 Artificial Intelligence Doing by program what could be assumed to require human intervention. Or I thought it was complicated but, now that I see your trick, it’s obviously a trivial task. Or ????

3 CSE5900 Lecture 9AI & MM Pt. 13 “The Cognitive Paradigm” To learn how humans perform mental tasks, write a program that replicates human performance (including errors, and (relative) solution times. To learn how to program a sophisticated task, find out how humans solve the problem and replicate in a program. Practice psychology by doing programming Practice programming by doing psychology (Except we call it cognitive science, not psychology…)

4 CSE5900 Lecture 9AI & MM Pt. 14 The Ebb and Flow of AI AI goes through periods of great popularity, followed by AI winters (little commercial interest) We are coming out of an AI winter right now, with neural network based data mining being the application domain of choice. Within academia, AI always is fascinating, and sometimes the source of grants. Note: I am an established AI skeptic, which to some extent is reflected in what follows.

5 CSE5900 Lecture 9AI & MM Pt. 15 Two AI Architectures Rule or Production systems Neural Networks or Fuzzy systems Production-based systems consist of –A large set of if-then statements (rule/production base) –These will usually have probabilities attached on both sides –A mechanism to march through the productions (inference engine) –The rule base is not ordered, except in use –The inference engine uses the common left and right hand terms to tie productions together. –The inference engine has in built problem-solving techniques, e.g., –SOAR-based systems use recursive sub-goaling.

6 CSE5900 Lecture 9AI & MM Pt. 16 Fuzzy Logic/ Neural Nets Consider a very large network of nodes. At the beginning, each node is weakly connected with all other nodes. An “example” is “presented”, which (somehow--who cares) increases some connections and decreases others. The network is “trained”. If the network provides the wrong answer (is this example a “cat”?) then the node connections that produced the answer are weakened, or the reverse. At the end of training, the network can usually tell cats from dogs, but there is no way of knowing how it does this or what traits are used how.

7 CSE5900 Lecture 9AI & MM Pt. 17 Two Different Usage Patterns Static –The system (rules, net, whatever) is prepared prior to use –Each running of the system varies only in the external events/characteristics encountered –Example: any diagnostic system Dynamic –The system changes (at least in the weighting) during use –The changed system may be saved and reused across runs. –Example: a game that improves as you improve

8 CSE5900 Lecture 9AI & MM Pt. 18 The Three Primary Areas of AI Use in MM Agents Games Computer Based Training (CBT, or CBI, or whatever)

9 CSE5900 Lecture 9AI & MM Pt. 19 Agents See N. Negroponte’s Being Digital for the fantasy. (Look carefully through the WWW for any working agent for the reality.) Agent: Someone to whom you delegate a task: Insurance agent, inquiry agent, confidential agent. The agent is capable of being trusted to do the delegated task. I might want to delegate: –Identifying and deleting incoming spam from my email. –Selecting news items I’ll probably want to read. –Doing comparison shopping –Keeping my larder stocked without my intervention

10 CSE5900 Lecture 9AI & MM Pt. 110 “AI Complete” (Like NP Complete) NP Complete is a concept from mathematics, referring to a class of problems where: –The algorithm to solve the problem is trivially obvious –But any non-trivial version will require an enormous (life of universe, plus) amount of time to run. –If you can figure out a fast way to solve one NPC problem, you’ve resolved all of them. –The Traveling Salesman Problem is the classic example –There is no way around the NPC

11 CSE5900 Lecture 9AI & MM Pt. 111 There Are Classes of AI Problems with Similar (Informal) Characteristics They seem trivial They can’t be solved in real time They are linked Probably if you solve one of them you solve all of them The Problems of Context and Common Sense are examples. Language (semantics, pragmatics) cannot be processed completely without a solution to these problems

12 CSE5900 Lecture 9AI & MM Pt. 112 Language Incomprehension Precludes Verbal (written or spoken) instructions Speech comprehension Realistic speech generation Handwriting recognition E-Mail filtering by agents News filtering by agents Unless –The system operates in a highly limited and constrained domain. Note: Games aren’t sufficiently constrained, so there are no longer text-based (typed, spoken) game interfaces

13 CSE5900 Lecture 9AI & MM Pt. 113 Then Why Do People Like Negroponte Keep On about Agents and Agency??? Why indeed….

14 CSE5900 Lecture 9AI & MM Pt. 114 Games In games, NPC = non-player character, that is, a character generated and controlled by program You don’t want to play against dumb opponents You do want your opponents’ skill levels to increase over time in line with your own. You do want your opponent’s tactics to reflect current situations. So we need lots of AI in games.


Download ppt "CSE5900 Lecture 9AI & MM Pt. 11 AI and MM Part 1 (An Embarrassingly Over-Simplified Introduction)"

Similar presentations


Ads by Google