1PAL © 2003 Carnegie Mellon University Space: The final frontier Crisis space/resource allocation LTI/CSD: Jaime Carbonell, Scott Fahlman, Eugene Fink LTI: Bob Frederking, Greg Jorstad, Ulas Bardak, Thuc Vu, Richard Wang Implicitly: Yiming Yang, William Cohen, Lori Levin, Steve Smith,… December 18, 2003
2PAL © 2003 Carnegie Mellon University People Jaime CarbonellScott Fahlman Bob FrederkingRichard Wang understanding Thuc VuUlas BardakGreg JorstadEugene Fink Space allocation
3PAL © 2003 Carnegie Mellon University Purpose Automated allocation of office space and related resources to office users, in both crisis and routine situations. Satisfy work-related needs of individual users and groups Maximize user satisfaction Ensure fair allocation of space
4PAL © 2003 Carnegie Mellon University Urgent space allocation Toxic Cloud! Toxic Cloud! - Alloc. solvable? - Decomposable? - Cope with surprise: not internet-wired, dispersed,… Wean Hall
5PAL © 2003 Carnegie Mellon University Main steps Elicitation of user preferences Near-optimal allocation based on partial knowledge of preferences Mediation of trades among users Negotiations with users
6PAL © 2003 Carnegie Mellon University Demo
7PAL © 2003 Carnegie Mellon University Main challenges How to represent and reason about space How to optimize space allocation conditioned on resources, constraints, preferences, and forecasts How to cope with uncertainty, such as partial knowledge of preferences, contingency planning based on possible exogenous events, and prediction of negotiation outcomes How to learn what works and why How to cope with surprise, such as crises, degraded space, new constraints, new preferences, new utility functions, and new optimization criteria
8PAL © 2003 Carnegie Mellon University Knowledge representation Need to represent: »Facts: People, departments, equipment, affinity groups,… »Spatial relations: 2D and 3D maps, connectivity, functions,… »Constraints: Minimal space/person, labs w/plumbing and power,… »Preferences: Proximity relations, windows, equipment,… »Utilities: Cost and benefit of satisfying preferences, elasticities,… »Episodes: Past space planning with utilities and outcomes, including justifications for decisions, retractions,… »Optimization criteria: As first-class objects, so as to reason about what to optimize and how to assign weights and priorities Need to reflect on: »What does not the system know that it needs to know: To trigger active learning, user interactions,… »What-if scenarios: For multi-user negotiation, to assess the completeness of the system’s knowledge,…
9PAL © 2003 Carnegie Mellon University Pervasive learning Learning at the factual level »By being told: Constraints, preferences, facts,… »By negotiation and examples: Preference weights, utilities,… Learning at the planning level »Mode selection: When to plan, when to seek information, when to negotiate, when to optimize, when to validate,… »Operator selection: How to select best actions within modes »Historical learning: What to reuse/transfer longitudinally Learning at the meta level »Self assessment: Utility of the learning (e.g. idiosyncratic versus general), accuracy of the learning, permanence,… »Targeting the learner: On maximizing expected future discounted utility, on correcting flaws (unlearning),…
10PAL © 2003 Carnegie Mellon University First fifteen months Non-crisis space allocation »Add users to an existing occupied building »Allocate offices in a new building Respecting constraints and preferences »Unary: Size, windows, internet, bio-isolation,… »N-ary: Proximity to co-workers, quiet,… Optimizing global utility »Maximize preference satisfaction »Minimize moving users already in place Coping with uncertainty »Use ranges, defaults, what-if planning »Elicit preferences and trade-offs »Support single-user and multi-user negotiations
11PAL © 2003 Carnegie Mellon University Architecture
12PAL © 2003 Carnegie Mellon University Main modules Natural-language communications Representation of user preferences, which includes defaults and learned knowledge Representation of (uncertain) knowledge of available space and related resources Optimization based on available knowledge Intelligent elicitation of user preferences and information about available space Single-user and multi-user negotiations Bartering office space among users Speed-up and quality learning
13PAL © 2003 Carnegie Mellon University Initial results Understanding of space-related Limited representation of space and related resources without uncertainty Optimization based on simple preferences
14PAL © 2003 Carnegie Mellon University Initial results: understanding Extraction System Template Generator Space Space Template Extraction Rules Extracted Text
15PAL © 2003 Carnegie Mellon University example Johnson wants to move to wean, he prefers the room He wants that room for conducting his experiments. His room will be filled with chemical bottles and equipment. He would like to be on the 5th floor, or higher than the 6th floor, but definitely not lower than the fourth floor please. He prefers the size of his room to be between 10–25 square meters. He will be moving into the room starting 2/28/2004 until May 24, He likes to have at least a window in his room, if possible. His room should have at least 2 doors. His room does not need internet, but definitely need electricity and 5–10 sinks. He would like to be above the Wean Engineering Library, and below his advisor's office. He would also like to be around 50 to 100 yards away from the building's entrance.
16PAL © 2003 Carnegie Mellon University Extraction rules Noun phrase identifier » A rule-based noun phrase chunker utilizing the part-of-speech tags from the tagger Name identifier » Identifies names; tolerant of uncased names » Compared with BBN IdentiFinder as baseline (IdentiFinder was trained on newswire text) » Evaluated on 124 s (test set); precision/recall based on entire name: 91%89% Our name identifier 65%54% BBN IdentiFinder PrecisionRecall
17PAL © 2003 Carnegie Mellon University Extraction rules Negative scope identifier » Determines what part of the sentence has negated meaning » … but not too far away from his classmates… Quantity identifier » Identifies quantities along with logical attributes » … must be a hundred fifty five square feet…
18PAL © 2003 Carnegie Mellon University Extracted text requester: Johnson filler: chemical bottles and equipment purpose: conducting his experiments building: wean room: 5102 date_start: 2/28/2004 date_end: May 24, 2004 floor_min: the 6th floor | the fourth floor floor: the 5th floor size_range: 10–25 square meters window_min: a window entrance_min: 2 doors entrance: the building's entrance plumbing_range: 5–10 internet_neg: internet electric: electricity rel_above: the Wean Engineering Library rel_beneath: his advisor's office dist3_range: 50 to 100 yards dist3_from: the building's entrance
19PAL © 2003 Carnegie Mellon University Space template after normalization [Request_Allocation requester: (#Johnson#) filler: (#chemical bottles and equipment#) purpose: (#conducting his experiments#) building: (WEH) room: (5102) floor: (5|>6|>4) size: (>108|<269) start_date: (Sat Feb ) end_date: (Mon May ) window: (+) entrance: (+|>2) internet: (-) plumbing: (+|>5|<10) electric: (+) above: (#the Wean Engineering Library#) beneath: (#his advisor's office#) distance: ((>150|<300)(#the building's entrance#)) ]
20PAL © 2003 Carnegie Mellon University Initial results: Representation List of available offices Database of basic office properties (size, windows, internet connections,…) On-demand computation of other properties (distance between offices, accessibility,…)
21PAL © 2003 Carnegie Mellon University Initial results: Conversion from AutoCad Identify the position of each office Compute the office areas Identify the office numbers Find the shortest paths between offices DB AutoCad RADAR/ Space Representation AutoCad maps include line drawings and free-floating text for office numbers.
22PAL © 2003 Carnegie Mellon University Transitions: » Assign a user to an office » Remove a user from an office » Exchange locations of two users Initial results: Optimization Application of simulated annealing. State: Assignment of users to offices; a user may be in a specific office or have no office Objective function: Weighted count of unsatisfied user preferences
23PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge » Uncertain knowledge of user preferences » Uncertain knowledge of available space » Allocation of space and related resources » Possible communications with users Use of Scone for knowledge representation and inference of implicit information
24PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge Understanding of space-related » Use of Scone knowledge base » Handling multiple requests in one » Handling user responses to earlier s » Identifying unclear places in s and asking users for clarification
25PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge Understanding of space-related Generation of replies » User-friendly explanations » Politeness and diplomacy
26PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge Understanding of space-related Generation of replies Optimization based on partial knowledge » Find an allocation with a (near-)largest expected value of the objective function » Estimate the standard deviation of the resulting expected value » Determine what additional information may reduce the standard deviation
27PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge Understanding of space-related Generation of replies Optimization based on partial knowledge Preference elicitation and negotiation » Select questions that reduce uncertainty » Estimate probabilities of possible replies » Reduce the number of s to users
28PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge Understanding of space-related Generation of replies Optimization based on partial knowledge Elicitation of preferences and negotiation Fairness of space allocation, with respect to » Novice users » Helpful users » Busy users
29PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge Understanding of space-related Generation of replies Optimization based on partial knowledge Elicitation of preferences and negotiation Fairness of space allocation Soft commitments and cancellations » Support different levels of commitment » If breaking a commitment to a user, negotiate appropriate compensation
30PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge Understanding of space-related Generation of replies Optimization based on partial knowledge Elicitation of preferences and negotiation Fairness of space allocation Soft commitments and cancellations Bartering among users » Allow users to offer office-space trades » Identify prospective multi-user trades
31PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge Understanding of space-related Generation of replies Optimization based on partial knowledge Elicitation of preferences and negotiation Fairness of space allocation Soft commitments and cancellations Bartering among users Interaction with human administrators » Providing relevant information » Asking help with complex decisions » Supporting multiple administrators
32PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge Understanding of space-related Generation of replies Optimization based on partial knowledge Elicitation of preferences and negotiation Fairness of space allocation Soft commitments and cancellations Bartering among users Interaction with human administrators Learning new knowledge and strategies » User preferences » Negotiation strategies » understanding
33PAL © 2003 Carnegie Mellon University Future tasks Representation of relevant knowledge Understanding of space-related Generation of replies Optimization based on partial knowledge Elicitation of preferences and negotiation Fairness of space allocation Soft commitments and cancellations Bartering among users Interaction with human administrators Learning new knowledge and strategies Graphical user interface » Visualization » Spatial input
34PAL © 2003 Carnegie Mellon University Schedule of initial versions Knowledge representation (March 2004) Understanding (March 2004) Generation of replies (March 2004) Partial-knowledge optimization (May 2004) Elicitation of preferences (May 2004) Fairness of space allocation (December 2004) Soft commitments and cancellations (2005) Bartering among users (2005) Interaction with human administrators (2005) Learning new knowledge (long term) Graphical user interface (long term)
35PAL © 2003 Carnegie Mellon University Interaction with other systems Scheduling » Determine the availability of specific users » Anticipate the needs of users and groups » Identify and prioritize space-related s » Estimate response times of specific users Webmaster » Provide on-line information about space User studies » Evaluate user satisfaction » Improve interaction with users
36PAL © 2003 Carnegie Mellon University Evaluation Allocation quality Speed and scalability Uncertainty tolerance Surprise tolerance User-friendliness
37PAL © 2003 Carnegie Mellon University To be continued…
38PAL © 2003 Carnegie Mellon University Allocation quality Maximizing quality of space allocation a i Subject to time constraints, available extrinsic data, available task information, and communication constraints Comparison with the results of omniscient unconstrained optimization a opt