SBIR MISSION-PLANNING

SBIR Proposal Summary on The Agency

The Lexikos Agency implements a three-part model of English language processing by its contained intelligent agents. Each such agent functionally integrates:

(A) The Lexikos Transcriber (now existing) as its textual input module;

(B) A web-based Generator for text paragraphs (and speech/GUI) outputs;

The Context module defines the logical DOMAIN (AKA microworld) the input and output paragraphs discuss. Each agent within The Agency gets its own evolving instance, which models and uses a hierarchy of partial-order PLANS featuring:

(D) At each node in the hierarchy, a MISSION, an AGENT (soldier, device, military unit…), and a dynamic STATE-SPACE (battle zone, supply depot, classroom…) in which the modeled AGENT seeks to achieve its MISSION.

(E) Verb-centric OPs (events and processes) modeled under Lexikos’ INGLISH ontology. Each OP cites expected STATE changes of its verb’s complements using PRECONDITION, EFFECT, and TENSE models. Each OP may be declared an AGENT-specific primitive, or decomposed into a new sub-PLAN.

(F) A PLANNING-LANGUAGE based on FOPC, defined in layers. All agents share the core layer, kept simple by design. A related planning algorithm tries to refine each partial PLAN using declared OPS plus “casual-link” sequencing constraints, aided by operator inputs. Optionally, real-time perceptions can be supported in extension layers. (Many are possible).

Transcriber exploits the PLAN-base as a flexible but robust model of “what is being discussed”. Generator uses it as AGENT to plan English outputs. The integration of Agency emphasizes shared usage of these other system elements:

(G) Vocabulary: STATE-SPACE literals, OP models, and other objects in Context help each agent link relevant meanings to TERMS in a core lexicon, which is otherwise too large. (The Agency’s base vocabulary holds 8,000 core roots plus Roget’s Thesaurus; each agent independently extends it.)

(H) Discourse model: a FOCUS mechanism tracks topics in Transcriber output charts, links them to PLAN elements, and relates them to abstract speech-acts via coherence relations. They tell the agent what each operator input sought to accomplish, so Generator can properly respond to it.

(I) Speaker and Listener models: an agent and the human operator typically play these fluent roles, but other bindings may occur inside PLANS. Good models of Speaker/Listener goals and speech acts are vital for coherent dialogs, and even to handle anaphora. Such models persist in Context.

(J) Evolving visual maps of Context: session-based web I/O specific to each dialog lets the operator monitor agent data structures to confirm that what her agent “understands” is indeed what she intended. This I/O channel also aids review of each PLAN as it develops or afterwards.

This functionality for The Agency is all packed into one portable Java Web App, to which operators may post English paragraphs and get replies. As each dialog progresses, a PLAN-base in FOPC evolves as a side effect, to be serialized as desired by the operator into metadata (RDF, etc.) that saves all recent changes.

Appendix 1: Static Extensions Now Planned

The theme of an agent planning a hierarchy of sequenced tasks within a limited domain has been studied in AI classrooms for decades. Combining it with English I/O is nearly as old, as I saw first from Winograd, the MIT teaching assistant in my ‘72 AI 1.01 course. His SHRDLU design got me “hooked” on AI and NLP.

Improved planning design patterns now exist. Each needs special-case handling, especially in a Parser/Planner hybrid like The Agency. They will extend its core PLANNING LANGUAGE – adding special-purpose English vocabulary and FOCP logic, plus related operator training. To help extensions rapidly develop, we plan an “open source” model for their software R&D. Targeted options include:

1) Complex end-states within the basic OP model: These involve competing core planning algorithms, each with specific refinements in goal decomposition, sub-task ordering, sequential constraints, conditional effects, disjunctive or negated goals, universal quantifiers, theorem-proving and other aspects of FOPC often mysterious to those not expert in logic programming.

2) Time and space models: expertise in them arises more often in NLP work, where verb tenses and prepositions need detailed exploration. Microworlds on both have developed able to support optional features for all OPs that need them, and pave the way for treating time and space as resources inside PLANs.

3) Uncertainty: planning strategies using coercion, abstraction, aggregation, and “intel” GOALS can reduce this effect. But for simple agent, any notion of uncertainty could be a flaw of overdesign, so we view it as an optional extension. Typical PLANs should come from agent reasoning, not guessing.

4) Belief networks: rules handling beliefs on (a)typical class members may be needed to support common NLP parsing heuristics. More complex extensions based on Bayesian reasoning, or other statistics-linked logics can probably co-exist in some PLANs, but will demand special training and vocabulary.

5) Utility Theory: complex preferences can be handled with models of money and economics; depletion of AGENT resources; and the costs of OPs based on simple formulae. Such extensions seem unneeded in the PLANs for most simple AGENTS. Complex AGENTS (e.g., teams) might use them, but details of such models seem likely to required AGENT-specific rules. Open source coding allows that.

Extensions such as (1-5) will augment or replace the core planning mechanisms of The Agency, so that more powerful types of PLANs can be discussed and refined by an operator tapping the needed training and specialized technical vocabulary.

Even with all the above, however, any PLAN emerging from such a dialog would be static – essentially frozen at design time, long before the AGENT who carries it out starts work on the details. Such plans may be helpful, but each can only be applied by passing the AGENT specific parameters, then watching it execute.

Interactions with the real world as AGENTS execute can make other types of PLAN more powerful. The Agency seeks also to support these dynamic PLANs, which can exploit real-time inputs and optimize their AGENTS’ behaviors accordingly.

Extensions that address such dynamic affects are discussed separately, as they require a PLANNING process more interleaved with PLAN execution, able to monitor what happens and/or change PLANs based on unexpected events in each environment, to more flexibly optimize overall AGENT performance.

Appendix 2: Dynamic Extensions for Active AGENT I/0

These extensions let a PLAN arise which incorporates future decisions made after the AGENT starts working on its tasks. They may be based on perceptions of each STATE-SPACE after it is accessible; or on bulk data coming in from other AGENTs under fixed a communication PLAN; or on new linguistic rules and info from any operator of The Agency with the rights to add them via Transcriber inputs.

This dynamic mode is what most people envision for “intelligent agents”. Key extensions to help handle it are below, simple and complex. Either way, they effectively make The Agency a smart interpreter for very high-level scripts:

1) Changeable AGENT Resources: with vocabulary for measures and fluents, each PLAN can include contingency sub-tasks chosen per real-time inputs. (Even Boolean inputs can help). E.g., “If you still have adequate resources, move on to your next planned task; else first solve the resource problem.”

2) Operator advice DURING execution (perhaps via PDA-like I/O channels): Such extensions let qualitative human judgments on STATES replace measurements. E.g., “The target structure appears empty; so now what do we do?”

3) Deferred sub-PLANs: reliably built DURING execution of a MISSION, these let many tasks be planned only when needed (versus exhaustively, all in advance). E.g. “Hey! I found one still operating. What should I do with it?”

4) Non-monotonic reasoning if assumptions later prove wrong. (Must be used with care, as standard FOPC rules may do not relate well). E.g, “Okay, I got to the radio as we planned, but it’s busted. So now what?”

5) Replanning based on new MISSIONS: they may adjust key elements, yet reuse most of a PLAN to save time and avoid starting from scratch. (E.g. “We must abort. The men need a quick exit or they’ll be trapped by that fire”.)

6) Metadata communication with similar AGENTS, arranged by passing bulk, low-level XML files. E.g. “Send unit #2 a dump of our current situation.”

7) Linguistic communication with similar AGENTS. E.g. “Ask unit #2 what help they can offer us on this unexpected problem.”

Important: such extensions are incremental, and must be. To create them, one first needs the static, lower-level parts of a Transcriber, a Generator, and a Context to be extended, plus a compatible, pre-constructed PLAN-base defining the MISSION-specific vocabulary, facts, and rules required to quickly absorb real-time data into the current Context and react to it semi-intelligently.

Without such a baseline, scenarios like those above – and the practical data- fusion and embedded-training scenarios they can jointly support – become wishful thinking and pipe dreams. If real MISSIONS are to include such scenarios, solid tools must arise first, able to help out on the bulk data entry of sub-PLANs.

The Agency will provide that help, by refactoring the Lexikos Transcriber to add partial parsing and web I/O; then adding a Context based on FOPC, planning and INGLISH verb models. We will then solicit extensions and a (voice) Generator to enhance the core, using help and advice from other AI experts and “open source” models of extension R&D, so that many smart people can semi-independently work in compatible ways. Please join us in refining and implementing this PLAN.