Modeling Human and Organizational Behavior Application to Military Simulations (1998) / Chapter Skim
Currently Skimming:
8 Planning
8 Planning
Pages 203-241
The Chapter Skim interface presents what we've algorithmically identified as the most significant single chunk of text within every page in the chapter.
Select key terms on the right to highlight them within pages of the chapter.
From page 203... ...
It then examines approaches taken to modeling planning first in military planning models, and then in the work of the artificial intelligence and behavioral science communities. The chapter concludes with conclusions and goals in the area of planning.
|
From page 204... ...
The mission analysis stage begins with receipt of an order from the unit's command and proceeds to more complete definition of the initial state (or, equivalently, current situation) , as well as a definition of the final goal state (or, equivalently, the mission objectives)
|
From page 205... ...
and planning activities (goal state specification)
|
From page 206... ...
· Planning strongly reflects the task context. Planning is Doctrinalized and Knowledge Intensive Because of the complexity of the domain, the risk/reward ratios of the outcome, the limited resources available for finite-time plan generation, and, perhaps most important, individual differences in planning and decision-making abilities, the planning process has been highly doctrinalized.3 As noted, step-bystep procedures are designed to take the planner from the current situation to the mission end state.
|
From page 207... ...
Planning at the higher echelons is more contemplative and is supported by extensive rational analyses of the options.5 This suggests that different planning techniques and processes may be employed by the tactical decision maker at each echelon. Planning at different echelons also affects how current situations (initial state)
|
From page 208... ...
In the artificial intelligence planning literature, this process is viewed as abstraction planning (Sacerdoti, 1974~. Planning Strongly Reflects the Resource Context Planning is not conducted in an abstract world of perfect and complete information (as in, for example, a chess game)
|
From page 209... ...
If the ultimate objective of future efforts in human behavior representation is to model doctrinally correct planning behavior, this should be an adequate starting point for a requirements definition effort. If, however, future efforts in human behavior representation are to be aimed at modeling actual tactical planning, a serious knowledge engineering effort is needed to build a behavioral database describing planning during combat decision making.
|
From page 210... ...
, information generators fail to present interpretations or implications of the briefed information, and finalized plans are often not presented to the commander for final review (Metlay et al., 1985~. Situation Assessment to Support Planning Assumptions Situation assessment for planning suffers from a number of shortcomings, including (Fallesen, 1993)
|
From page 211... ...
is rejected the decision maker either moves on to a totally different option or idea or goes back up the deepening chain to a point (theoretically) above the source of the flaw and then follows another branch.
|
From page 212... ...
Clearly, one can expect the real-time replanning effort to deviate significantly from the doctrinally specified preoperations planning process. Implications for Modeling the Tactical Planning Process The above discussion has a number of key implications for modeling the tactical planning process, related to the difference between theory and practice; the need for domain knowledge in model development; the need to model the various stages of planning; and the effects of echelon, resources, and context.
|
From page 213... ...
NOTE: HBR, human behavior representations. training area, one might take the upper left option (doctrinal, perfect)
|
From page 214... ...
A decision-theoretic approach might be taken for doctrinally correct planning, whereas a recognition-primed decision making approach, relying on episodic memory, might be better suited to the modeling of actual planning behaviors. Whichever approach is taken, attention needs to be paid to whether an optimizing or a satisficing approach is used.
|
From page 215... ...
MODELS FOR PLANNING IN MILITARY HUMAN BEHAVIOR REPRESENTATIONS The following three subsections (1) review planning models in existing military human behavior representations; (2)
|
From page 216... ...
human behavior representation development effort. It also identifies the key points of interaction among team members (e.g., "provide commander's guidance to staff," "coordinate with division G2")
|
From page 217... ...
environment. The CGF plan generation module consists of 22 separate modules or rulesets, each containing 10 to 100 production rules, which encode a subject matter expert's ranking of decision alternatives for a number of key decisions, such as attack or abort, avoid contact, select assault point, and select route.
|
From page 218... ...
Marine Defense Advanced Marine (HBR models) Individual Route planning Computer- Research Projects and unit Generated Force Agency (DARPA)
|
From page 219... ...
PLANNING ~ntations 219 Functions/ Implementation/ Activities Modeled Architecture Comments ividual and Route planning · Genetic algorithm optimization · Currently an exploratory study lade Course-of-action · Decision-aiding workstation with · Underlying command/staff generation graphic depiction of military dependency matrix provides over entities all framework for information flow among HER agents ividual Route planning · Production rule system driven by · Multiple rulesets for different unit external state variables decision alternatives · Generation of single plan frame for ModSAF execution ividual Short-term planning · Hierarchical goal decomposition · Appears to be single-step for individual · Decision-theoretic goal selection planner, but could be expanded military operations in · Situation-driven rules to multiple steps for individual urban terrain course-of-action generation activities ividual and Full activities of · Soar architecture with · Planning not explicitly ; tactical pilots across hierarchical goal decomposition represented, as Soar supports a wide range of · Efficient production rule system only single-step planning aircraft and missions to deal with large rulebase · Could be expanded to support · Situation-driven rules an explicit planning module ividual and Full activities of · Live battalion commander · Plan generation and ; (company) rotorcraft pilots and · Soar-CFOR company commander elaboration done through company commander that: tactical templates and standard for RWA mission Generates mission plan operating procedures Monitors progress · Plan refinement done through Replans as necessary checks on task · Soar-IFOR RWA pilots interdependencies, timing · ModSAF vehicle entities ividual and Full activities of · Symbolic operator model · Similar to Soar in its top-down ; tactical rotorcraft architecture with hierarchical decomposition from mission pilots mission activity decomposition phases to low-level activities · Production rule system to · Single-step planning, but implement procedures could be expanded to support · Situation-driven productions an explicit planning module continued
|
From page 220... ...
Most of the activity appears to be fairly reflexive, but there is an attempt at situation-driven planning using hierarchical goal decomposition. High-level goals 10 Formerly known as the team target engagement simulator (TTES)
|
From page 221... ...
AGW planning Course-of-action generation Course-of-action generation and logistics planning · Decision tables or prioritized production rules · Future state predictor that supports some planning Four stage process: 1. Terrain analysis for route planning 2.
|
From page 222... ...
-97 large-scale warfighting simulation (Laird, 1996~. The standard Soar approach to hierarchical goal decomposition is taken, so that a high-level goal (e.g., "conduct intercept")
|
From page 223... ...
In generating a path through the goal tree and subsequently pruning off lower-priority paths, Soar has the capability of generating a linear sequence of goals and subgoals to be followed over some finite time horizon, starting with the current time frame. This linear sequence of goals and their associated behaviors is effectively a plan: it is a linear sequence of behaviors that will transform the current state of the system into some desired goal state through a path of sequential system states defined by the subgoals along the way.
|
From page 224... ...
The planning component uses an approach based on constraint satisfaction to link together component activities, which are doctrinally specified, and which, when fully linked, form an activity plan, taking the command entity from the current situation to the desired end state. Specific details of how activities are linked efficiently to avoid a combinatorial explosion of options through the planning space are described in Calder et al.
|
From page 225... ...
If the two do not match sufficiently closely, replanning is triggered. A production rule framework is used to check for antecedent satisfaction for plan monitoring and to maintain a memory of the plan goal for subsequent backward chaining and replanning.
|
From page 226... ...
, such as relying on internal perceptions of external states, but it is this common approach to goal/procedure decomposition that suggests, at least in the planning realm, that MIDAS and Soar are effectively equivalent representational approaches: Both are reflexive production rule systems,~4 in which the production rules representing behaviors at various levels of abstraction are carefully crafted to generate the appropriate behavioral time/event sequences for the scenarios and environment for which they were designed. Both are one-step or single-frame planners, but, as discussed earlier for Soar, both could be transitioned to multiframe planners, apparently without great difficulty.
|
From page 227... ...
Thus instead of "do action a with priority p," we might see "generate plan a with priority p." However, it is unclear whether actual plan generation is triggered by this production rule, or prestored contingency plans are retrieved from memory and issued by the command entity. An intermediate possibility, however, is suggested by the existence of another NSS module, the future enemy state predictor.
|
From page 228... ...
It would appear that this type of simulation-based evaluation should be a key component of any planning model considered for human behavior representation. A concurrent effort at the Institute for Simulation and Training is the development of a unit route planning module for modeling route planning at multiple echelons (battalion, company, platoon)
|
From page 229... ...
. The BRS is intended to serve as a prototype decision aid at the battalion through corps echelons, but two aspects of the system suggest that it may have potential for the development of future models for human behavior representation: (1)
|
From page 230... ...
If so, it still remains to be seen whether this approach maintains some degree of psychological validity with respect to what is seen with the operations staff. Decision Support Display The effort to develop the decision support display (DSD)
|
From page 231... ...
However, the DSD knowledge base may be useful in other applications, particularly planning models for human behavior representation, if the rulebase effectively codifies current tactical guidelines, rules, and decision heuristics. As this is an ongoing project under the auspices of the Federated Laboratory program, it should be rereviewed again at a later date.
|
From page 232... ...
The review of military planning models in this section reinforces a key finding of the previous section on tactical decision making: military planning is extremely knowledge intensive. Although early artificial intelligence planners emphasized general problem-solving capabilities in an effort to develop domainindependent "planning engines," it seems clear that the military planning domain, especially Army planning, requires that planning be conducted in accordance with an extensive doctrine covering operations and intelligence at all echelons in the military hierarchy.
|
From page 233... ...
This is the function of "upstream" perceptual and situation assessment submodels, which are represented at varying levels of fidelity in the models reviewed here: · In some cases, neither the perceptual nor situation assessment functions are explicitly represented (e.g., FWA-Soar) , so that environmental state variables (e.g., the identity and location of a bogey)
|
From page 234... ...
Modeling approaches for dealing with multitasking of this sort are discussed further in Chapter 4, and the findings presented there are clearly appropriate to the modeling of in-context military planning. PLANNING MODELS IN THE ARTIFICIAL INTELLIGENCE AND BEHAVIORAL SCIENCE COMMUNITIES An extensive literature on planning models exists in the artificial and behavioral science communities, and it is beyond the scope of this study to do much more than provide the reader with pointers to the relevant literature.
|
From page 235... ...
Plan goals/tasks can have conditionals. Plans can have conditions for task initiation, termination, or branching, and there are several ways of dealing with this issue.
|
From page 236... ...
A slightly more recent review of artificial intelligence planners, again cover ing approximately 150 citations, is provided by Akyurek (1992) , who proposes the basic planner taxonomy shown in Table 8.5.
|
From page 237... ...
observations on how current planners are attempting to deal with these problems. TABLE 8.6 Advanced Artificial Intelligence Planner Techniques World Characteristics Planner Strategy Fully observable, static, deterministic Partially observable Dynamic Stochastic Use classical planners Specify information-gathering strategy Interleave planning and acting Use Markov models A more recent although not as extensive review of cognitive architectures that includes planning functions is to be found at the following University of Michigan website: CapabilLists/Plan.html.
|
From page 238... ...
One last key shortcoming of artificial intelligence planners in general is due to the artificial intelligence planning community's interest in developing computationally effective planners, and not necessarily in developing models of human planners. Thus, most artificial intelligence specialists working on planner development efforts pay scant attention to how actual human planners solve problems.
|
From page 239... ...
. They found, however, that this model fails to capture the planning behaviors they observed.
|
From page 240... ...
It is not clear that any particular community has the answer to the development of sophisticated planning models, algorithms, or decision aids, especially at the higher conceptual levels of plan generation (although there are many domain-specific tools to support the more tedious plan elaboration tasks)
|
From page 241... ...
· Develop planning models that are more reactive to account for plan failures, dynamically changing environments, and changes in plan goals. The focus should be on developing planners that are more robust to failures, develop contingency plans, and can rapidly replan on the fly.
|
Key Terms
This material may be derived from roughly machine-read images, and so is provided only to facilitate research.
More
information on Chapter Skim is available.