- Количество слайдов: 30
Human-Aware AI (aka Darned Humans: Can’t Live with them. Can’t Live without them) Subbarao Kambhampati Arizona State University Given at U. Washington on 11/2/2007
51 year old field of unknown gender; Birth date unclear; Mother unknown; Many purported fathers; Constantly confuses holy grail with daily progress; Morose disposition
So let’s see if the future is going in quite the right direction…
What is missing in this picture? (Static vs. Dynamic) (Observable vs. Partially Observable) Goals on (Full vs. Partial satisfaction) ac ti (perfect vs. Imperfect) perception Environment (Instantaneous vs. Durative) (Deterministic vs. Stochastic) What action next? A: A Unified Brand-name-Free Introduction to Planning T he $$ $$ $ tio s ue Q $ Subbarao Kambhampati
AI’s Curious Ambivalence to humans. . • Our systems seem happiest – either far away from humans – or in an adversarial stance with humans You want to help humanity, it is the people that you just can’t stand…
What happened to Co-existence? • Whither Mc. Carthy’s advice taker? • . . or Janet Kolodner’s house wife? • …or even Dave’s HAL? • (with hopefully a less sinister voice)
Why aren’t we doing HAAI? • . . to some extent we are – Assistive technology; Intelligent user interfaces; Augmented cognition, Human-Robot Interaction – But it is mostly smuggled under the radar. . – And certainly doesn’t get no respect. . • Rodney Dangerfield of AI? • Is it time to bring it to the center stage? – Having them as applied side of AI makes them seem peripheral, and little formal attention gets paid to them by the “main-stream”
(Some) Challenges of HAAI – Human-level communication/interfacing • Need to understand what makes natural interfaces. . – Explanations • Humans want explanations (even if fake. . ) • Teachability – Advice Taking (without lobotomy) • Elaboration tolerance – Dealing with evolving models • You rarely tell everything at once to your secretary. . – Need to operate in an “any-knowledge” mode • Recognizing Human’s state – Recognizing intent; activity – Detecting/handling emotions/affect Human-aware AI may necessitate acting human (which is not necessary for non-HAAI) • Communication
Caveats & Worries about HAAI • Are any of these challenges really new? • HAAI vs. HDAI (human-dependent AI) – Human dependent AI can be enormously lucrative if you find the right sweet spot. . • But will it hamper eventual progress to (HA)AI? • Advice taking can degenerate to adviceneeding. . • Designing HAAI agents may need competence beyond computer science. .
Are the challenges really new? Are they too hard? • Isn’t any kind of feedback “advice giving”? Isn’t reinforcement learning already foregrounding “evolving domain models” – A question of granularity. There is no need to keep the interactions mono-syllabic. . • Won’t communication require NLP and thus become AIcomplete? – There could well be a spectrum of communication modalities that could be tried • Doesn’t recognition of human activity/emotional state really AI? –. . it is if we want HAAI (you want to work with humans, you need to have some idea of their state. . )
HDAI: Finding“Sweet Spots” in computer-mediated cooperative work • It is possible to get by with techniques blithely ignorant of semantics, when you have humans in the loop – All you need is to find the right sweet spot, where the computer plays a pre-processing role and presents “potential solutions” – …and the human very gratefully does the in-depth analysis on those few potential solutions • Examples: – The incredible success of “Bag of Words” model! • Bag of letters would be a disaster ; -) • Bag of sentences and/or NLP would be good –. . but only to your discriminating and irascible searchers ; -) • Concern: – Will pursuit of HDAI inhibit progress towards eventual AI? • By inducing perpetual dependence on (rather than awareness of) the human in the loop?
Delusions of Advice Taking: Give me Advice that I can easily use • Planners that expect “advice” that is expressed in terms of their internal choice points – HSTS, a NASA planner, depended on this type of knowledge. . • Learners that expect “advice” that can be easily included into their current algorithm – “Must link”/ “Must-not Link” constraints used in “semi-supervised” clustering algorithms Moral: It is wishful to expect advice that will be tailored to your program internals. Operationalizing high-level advice is your (AI program’s) responsibility
HAAI pushes us beyond CS… • By dubbing “acting rational” as the definition of AI, we carefully separated the AI enterprise from “psychology”, “cognitive science” etc. • But pursuit of HAAI pushes us right back into these disciplines (and more) – Making an interface that improves interaction with humans requires understanding of human psychology. . • E. g. studies showing how programs that have even a rudimentary understanding of human emotions fare much better in interactions with humans • Are we ready to do HAAI despite this push beyond comfort zone?
How are sub-areas doing on HAAI? I’ll focus on “teachability” aspect in two areas that I know something about • Automated Planning – Full autonomy through complete domain models – Can take prior knowledge in the form of • Domain physics • Control knowledge –. . but seems to need it • Machine Learning. . – Full autonomy through tabula rasa learning over gazillion samples – Seems incapable of taking much prior knowledge • Unless sneaked in through features and kernels. .
What’s Rao doing in HAAI? • Model-lite planning • Planning in HRI scenarios • Human-aware information integration
Motivations for Model-lite Is the only way to get more applications is to tackle more and more expressive domains? • There are many scenarios where domain modeling is the biggest obstacle – Web Service Composition • Most services have very little formal models attached – Workflow management • Most workflows are provided with little information about underlying causal models – Learning to plan from demonstrations • We will have to contend with incomplete and evolving domain models. . • . . but our approaches assume complete and correct models. .
From “Any Time” to “Any Model” Planning Model-Lite Planning is Planning with incomplete models • . . “incomplete” “not enough domain knowledge to verify correctness/optimality” • How incomplete is incomplete? • Knowing no more than I/O types? • Missing a couple of preconditions/effects or user preferences?
Challenges in Realizing Model-Lite Planning 1. Planning support for shallow domain models [ICAC 2005] 2. Plan creation with approximate domain models [IJCAI 2007, ICAPS Wkshp 2007] 3. Learning to improve completeness of domain models [ICAPS Wkshp 2007]
Challenge: Planning Support for Shallow Domain Models • Provide planning support that exploits the shallow model available • Idea: Explore wider variety of domain knowledge that can either be easily specified interactively or learned/mined. E. g. • I/O type specifications (e. g. Woogle) • Task Dependencies (e. g. workflow specifications) – Qn: Can these be compiled down to a common substrate? • Types of planning support that can be provided with such knowledge – Critiquing plans in mixed-initiative scenarios – Detecting incorrectness (as against verifying correctness)
Challenge: Plan Creation with Approximate Domain Models • Support plan creation despite missing details in the model. The missing details may be (1) action models (2) cost/utility models • Example: Generate robust “line” plans in the face of incompleteness of action description – View model incompleteness as a form of uncertainty (e. g. work by Amir et. al. ) • Example: Generate Diverse/Multi-option plans in the face of incompleteness of cost model – Our IJCAI-2007 work can be viewed as being motivated this way. . Note: Model-lite planning aims to reduce the modeling burden; the planning itself may actually be harder
Imprecise Intent & Diversity
Challenge: Learning to Improve Completeness of Domain Models • In traditional “model-intensive” planning learning is mostly motivated for speedup –. . and it has gradually become less and less important with the advent of fast heuristic planners • In model-lite planning, learning (also) helps in model acquisition and model refinement. – Learning from a variety of sources • Textual descriptions; plan traces; expert demonstrations – Learning in the presence of background knowledge • The current model serves as background knowledge for additional refinements for learning • Example efforts – Much of DARPA IL program (including our LSP system); PLOW etc. – Stochastic Explanation-based Learning (ICAPS 2007 wkhop) Make planning Model-lite Make learning knowledge (model) rich
Learning & Planning with incomplete models: A proposal. . Represent incomplete domain with (relational) probabilistic logic – Weighted precondition axiom – Weighted effect axiom – Weighted static property axiom • Address learning and planning problem – Learning involves • Updating the prior weights on the axioms • Finding new axioms – Planning involves • Probabilistic planning in the presence of precondition uncertainty • Consider using Max. Sat to solve problems in the proposed formulation DARPA Integrated Learning Project • Towards Model-lite Planning - Sungwook Yoon
Indiana Univ; ASU Stanford, Notre Dame MURI 2007: Effective Human-Robot Interaction under Time Pressure Minh Do - PARC Planning with Goal Utility Dependencies
[CIDR 07; VLDB 07] Challenges in Querying Autonomous Databases Imprecise Queries Incomplete Data User’s needs are not clearly defined hence: Databases are often populated by: § § Queries may be too general Queries may be too specific § § Lay users entering data Automated extraction Density Function Relevance Function General Solution: “Expected Relevance Ranking” Challenge: Automated & Non-intrusive assessment of Relevance and Density functions However, how can we retrieve similar/ incomplete tuples in the first place? Challenge: Rewriting a user’s query to retrieve highly relevant Similar/ Incomplete tuples Once the similar/incomplete tuples have been retrieved, why should users believe them? Challenge: Provide explanations for the uncertain answers in order to gain the user’s trust QUIC: Handling Query Imprecision & Data Incompleteness in Autonomous Databases
Summary: Say Hi to HAAI • We may want to take HAAI as seriously as we take autonomous agency – My argument is not that everybody should do it, but rather that it should be seen as “main stream” rather than as some applied • HAAI does emphasize specific technical challenges: Communication; Teachability; Human state recognition • Pursuit of HAAI involves pitfalls (e. g. need to differentiate HDAI and HAAI) as well as a broadening of focus (e. g. need to take interface issues seriously) • Some steps towards HAAI in planning
Points to Ponder. . • Do we (you) agree that we might need human-aware AI? • Do you think anything needs to change in your current area of interest as a consequence? • (What)(Are there) foundational problems in human-aware AI? – Is HAAI moot without full NLP? • How do we make progress towards HAAI – Is IUI considered progress towards HAAI? – Is model-lite planning? – Is learning by X (X= “demonstrations”; “being told”…)? – Is elicitation of utility models/recognition of intent?
Epilogue: HAAI is Hard but Needed. . • The challenges posed by HAAI may take us out of the carefully circumscribed goals of AI • Given a choice, us computer scientists would rather not think about messy human interactions. . • But, do we really have a choice?