How to AI COGS 105. Traditional Rule Concept. if (wus=="hi") { was = "hi back to ya"; }

Transcription

1 COGS 105 Week 14b: AI and Robotics How to AI Many robotics and engineering problems work from a taskbased perspective (see competing traditions from last class). What is your task? What are the inputs and outputs to your agent? What are the rules that lead from input to output? Traditional Rule Concept In many respects, classic AI started with the concept of a production rule. An if-then rule that determines what the agent does if some condition holds. Example if-then rule in JavaScript: == compares what user said (wus) to hi if (wus=="hi") { was = "hi back to ya"; } Simon & Newell, 50 s Production rules can be more complicated by using patternmatching (and pattern-generating) algorithms. = sets what it says (was) to hi back...

2 Other pattern matches... indexof = check to see if the word mother appears in what they said. if (wus.indexof("mother")>=0) {... if (wus=="hi" wus=="hello" wus=="hey" wus=="hiya") {... if (wus.length > 10) {... length = how long, in characters, is what they said? = or, to check multiple possible patterns Where Do Rules Come From? Imagine trying to build an artificial agent that can engage in conversations. What does the agent need to know? Words How to line up words into sentences How to choose which sentence topics to pursue How to stay on topic How to guide the flow of conversation (e.g., question answer) Where Do Rules Come From? Imagine trying to build an artificial agent that can engage in conversations. Where Do Rules Come From? What does the agent need to know? Morphology Syntax Semantics Pragmatics / Discourse Pragmatics / Discourse

3 Conversation Analysis SEMAINE Project The analysis of interaction using video and audio, and then careful hand-coding and observation. Qualitative method. Many of the methods in conversation analysis can be transformed, in simple ways, into production rules. For example, the adjacency pair. E.g., offer/refusal (or acceptance), questionanswer, compliment-response Example: USC ICT Projects Another: SimCoach USC s Institute for Creative Technologies They develop sophisticated intelligent virtual characters that can engage in interaction using sets of sophisticated rules.

4 Cutting Edge: Authoring Tools ELITE The field of virtual agents is especially active in developing and testing authoring tools, to make it easier for researchers or people in industry to create artificial agents. RoundTable These tools have a point and click strategy, and allow users to build large repositories of patternmatching algorithms (production rules, with some spice) without having to program.

5 demo from previous year (just for illustration) demo from previous year (just for illustration) A Simple Interpreter Designed for this Purpose demo from previous year (just for illustration)

6 Two AI Cultures cognaction.org/ cogs105/chatagent GOFAI-based Embodied Dynamic Situated Hybrid Systems Robotics / AI The field is massive. Even just the tools built into this artificial, virtual software agent exemplify the array of domains of this field. action modeling HCI speech recognition speech synthesis computational linguistics Various Issues in AI, Etc. One problem faced by engineers and scientists is that robotic systems are becoming complicated enough that it is difficult to do on board. Solution: The cloud. From reading #2

7 Benefits Increased computational power off board : using high-performance servers to process / crunch numbers off the system itself. Increased distribution of robotic (or AI) agents can solve a problem faster, in parallel. Involvement of CogSci If you know how to code, you can be involved in all stages machine learning, neural networks, computer vision, etc. If you do not or do not want to code, you can still be involved by designing usability studies, serving as a domain expert to help coders, analyze human behavior, etc. From reading #2 Phamduy et al. Project Ideas I can open up the chatbot system for you to create your own chat agent. You can use the Braitenberg system to do some exploration of agent-based simulations (extending the last lab). with children You can do a usability study (ask me for readings, explaining how to do it).