1 ai mechanisms
1.1 symbolic approaches
1.2 distributed approaches
1.3 dynamic planning approaches
1.4 others
ai mechanisms
generally, artificial action selection mechanisms can divided several categories: symbol-based systems known classical planning, distributed solutions, , reactive or dynamic planning. approaches not fall neatly 1 of these categories. others more providing scientific models practical ai control; these last described further in next section.
symbolic approaches
early in history of artificial intelligence, assumed best way agent choose next compute optimal plan, , execute plan. led physical symbol system hypothesis, physical agent can manipulate symbols necessary , sufficient intelligence. many software agents still use approach action selection. requires describing sensor readings, world, of ones actions , of 1 s goals in form of predicate logic. critics of approach complain slow real-time planning , that, despite proofs, still unlikely produce optimal plans because reducing descriptions of reality logic process prone errors.
satisficing decision-making strategy attempts meet criteria adequacy, rather identify optimal solution. satisficing strategy may often, in fact, (near) optimal if costs of decision-making process itself, such cost of obtaining complete information, considered in outcome calculus.
goal driven architectures – in these symbolic architectures, agent s behaviour typically described set of goals. each goal can achieved process or activity, described prescripted plan. agent must decide process carry on accomplish given goal. plan can expand subgoals, makes process recursive. technically, more or less, plans exploits condition-rules. these architectures reactive or hybrid. classical examples of goal driven architectures implementable refinements of belief-desire-intention architecture jam or ive.
excalibur research project led alexander nareyek featuring any-time planning agents computer games. architecture based on structural constraint satisfaction, advanced artificial intelligence technique.
distributed approaches
in contrast symbolic approach, distributed systems of action selection have no 1 box in agent decides next action. @ least in idealized form, distributed systems have many modules running in parallel , determining best action based on local expertise. in these idealized systems, overall coherence expected emerge somehow, possibly through careful design of interacting components. approach inspired neural networks research. in practice, there centralised system determining module active or has salience. there evidence real biological brains have such executive decision systems evaluate of competing systems deserves attention, or more properly, has desired actions disinhibited.
asmo attention-based architecture developed rony novianto. orchestrates diversity of modular distributed processes can use own representations , techniques perceive environment, process information, plan actions , propose actions perform.
various types of winner-take-all architectures, in single selected action takes full control of motor system
spreading activation including maes nets (ana)
extended rosenblatt & payton spreading activation architecture developed toby tyrrell in 1993. agent s behaviour stored in form of hierarchical connectionism network, tyrrell named free-flow hierarchy. exploited example de sevin & thalmann (2005) or kadleček (2001).
behavior based ai, response slow speed of robots using symbolic action selection techniques. in form, separate modules respond different stimuli , generate own responses. in original form, subsumption architecture, these consisted of different layers monitor , suppress each other s inputs , outputs.
creatures virtual pets computer game driven three-layered neural network, adaptive. mechanism reactive since network @ every time step determines task has performed pet. network described in paper of grand et al. (1997) , in creatures developer resources. see creatures wiki.
dynamic planning approaches
because purely distributed systems difficult construct, many researchers have turned using explicit hard-coded plans determine priorities of system.
dynamic or reactive planning methods compute 1 next action in every instant based on current context , pre-scripted plans. in contrast classical planning methods, reactive or dynamic approaches not suffer combinatorial explosion. on other hand, seen rigid considered strong ai, since plans coded in advance. @ same time, natural intelligence can rigid in contexts although fluid , able adapt in others.
example dynamic planning mechanisms include:
finite-state machines these reactive architectures used computer game agents, in particular first-person shooters bots, or virtual movie actors. typically, state-machines hierarchical. concrete game examples, see halo 2 bots paper damian isla (2005) or master s thesis quake iii bots jan paul van waveren (2001). movie example, see softimage.
other structured reactive plans tend little more conventional plans, ways represent hierarchical , sequential structure. some, such prs s acts , have support partial plans. many agent architectures mid-1990s included such plans middle layer provided organization low-level behavior modules while being directed higher level real-time planner. despite supposed interoperability automated planners, structured reactive plans hand coded (bryson 2001, ch. 3). examples of structured reactive plans include james firby s rap system , nils nilsson s teleo-reactive plans. prs, raps & trp no longer developed or supported. 1 still-active (as of 2006) descendent of approach parallel-rooted ordered slip-stack hierarchical (or posh) action selection system, part of joanna bryson s behaviour oriented design.
sometimes attempt address perceived inflexibility of dynamic planning, hybrid techniques used. in these, more conventional ai planning system searches new plans when agent has spare time, , updates dynamic plan library when finds solutions. important aspect of such system when agent needs select action, solution exists can used (see further anytime algorithm).
others
cognitao decision making engine based on bdi (belief-desire-intention), includes built in teamwork capabilities.
soar symbolic cognitive architecture. based on condition-action rules known productions. programmers can use soar development toolkit building both reactive , planning agents, or compromise between these 2 extremes.
act-r similar soar. includes bayesian learning system prioritize productions.
abl/hap
fuzzy architectures fuzzy approach in action selection produces more smooth behaviour can produced architectures exploiting boolean condition-action rules (like soar or posh). these architectures reactive , symbolic.
^ karen l. myers. prs-cl: procedural reasoning system . artificial intelligence center. sri international. retrieved 2013-06-13.
Comments
Post a Comment