Abstract agent architectures : standard agent. Abstract agent architectures. Several types of agents (2) Several types of agents. Agent architectures

Transcription

1 Abstract agent architectures : standard agent n S = {s 1, s 2, } environment states n A = {a 1, a 2, } s n Abstractly, an agent can be viewed as a function: : S* Æ A mapping sequences of environments to s Abstract agent architectures n Nondeterministic behavior environment: env: S A Æ P(S) n History: inter agent and environment: a h: s 0 Æ 0 a s 1 Æ 1 a s 2 Æ Several types of agents n Behavior of a purely reactive agent: : S Æ A n Agent with perception: see: S Æ P : P* Æ A Several types of agents (2) n Agents with a(n internal) state: see: S Æ P : I Æ A next: I P Æ I see next agent state 190 environment 191 Several types of agents (3) Agent architectures n Deliberative agent see: S Æ P : DB Æ A next: DB P Æ DB see next agent DB environment 192 n Deliberative systems Based on physical symbol system hypoth.: GOFAI approach n Reactive systems Based on physical grounding hypothesis n Hybrid systems Synthesis of deliberative and reactive systems 193 1

2 Deliberative systems BDI Architecture / PRS n Uses symbolic representations, e.g.: Maintains s, e.g. world Planning (e.g. STRIPS) (Logical) Reasoning Sense-plan/reason-act cycle Belief, desires, intentions (BDI) Internal ( mental ) state (belief base, ) Beliefs Interpreter Sensor input Goals Plan library Intentions IRMA (Bratman et al. 1987) BDI architecture perception plan library means-end reasoner beliefs reasoner intention options intention structure filtering process surviving options deliberation process opportunity analyzer options desires 196 Sensor input belief revision beliefs generate options desires filter intentions 197 Abstract BDI Interpreter (Rao & Georgeff 1991) n initialize-state(); n do options := option-generator(eventqueue,b,g,i); selected-options := deliberate(options,b,g,i); update-intentions(selected-options,i); execute(i); get-new-external-events(); drop-successful-attitudes(b,g,i); drop-impossible-attitudes(b,g,i); n until quit. 198 Problems with deliberative systems n Representation & symbol grounding problem How to translate the real world into an accurate & adequate symbolic description How do symbols relate to real-world entities and processes? n Reasoning is computation-intensive How to get agents to reason with real world information in time for the results to be useful 199 2

3 Reactive systems Subsumption architecture (Brooks 1991) Problems with deliberative systems fi n `Re : reactive systems n Stimulus-response approach n No symbolic reasoning, no search n No internal, symbolic representations n Layered / hierarchical architecture sensing explore wander acting avoid obstacles Subsumption architecture: (hard-wired) control by suppression and inhibition n S: signal suppresses (and replaces) input into module for some predetermined time n I: signal inhibits output from module for some pre-determined time inhibitor Hybrid systems n Trying to synthesize both reactive and deliberative systems n Typically (levels Æ) layered: either vertically or horizontally S Module I suppressor Levels in hybrid systems Shakey architecture (Nilsson 1984) n Levels, e.g. three-level architecture: Low-level: reflexes (e.g. servo control) Intermediate level: combines low-level s into more complex behaviors High-level: planning (e.g. STRIPS planner) n Classical example: Shakey 204 Executor Visual perception Sensory signals STRIPS rules STRIPS planner Models Bump and touch detectors Plan generalizer Plan library Intermediate level s Low-level s Motor commands 205 3

4 Layered hybrid systems n horizontally or vertically layered Layered hybrid systems (ctd) Layer n Layer n Perceptual input Layer n Layer 2 Layer 1 Action output Layer 2 Layer 1 Layer 2 Layer 1 Perceptual input Perceptual.input Horizontal layering Vertical layering (1 pass control) Vertical layering (2 pass control) Hierarchical Robotic System Architecture (vertically layered) Example: Delivery Robot ROBOT DELIVERY ROBOT plan Controller-n Controller-2 Controller-1 arrived Follow plan Go to location & avoid obstacles goal_pos Percepts body Commands robot_pos Steer robot & report obstacles / position steer Stimuli Actions ENVIRONMENT ENVIRONMENT (Concrete) hybrid architectures TouringMachine (Ferguson 1992) n TouringMachine (horizontally layered) n InteRRaP (vertically layered)

5 InteRRaP (Müller 1996) InteRRaP: agent control unit Agent KB Social Mental World Agent Control Unit Cooperative Planning layer Local Planning layer Behavior-based layer Co-operative Planning Layer (CPL) Local Planning Layer (LPL) Behaviour-Based Layer (BBL) World Interface (WIF) perception communication 212 E N V I R O N M E N T 213 InteRRaP generic layer Kinds of hybrid systems (Arkin) Knowledge Base situation recognition execution Layer Control unit goal plan activation selection scheduling n Role of deliberation/planning in system: Selecting lower-level s/behaviors Adapting lower-level s/behaviors (smaller but more frequent modifications) Advising lower-level s (planner output competes with the lower-level s) Postponing reactivity by using the planner itself for generating res Towers in hybrid systems Triple-Tower Architecture n In hybrid systems there are hierarchies both on the perception and the side (called towers, Albus) n Nilsson observed that there is also a hierarchy in the part: a third tower, thus obtaining the so-called Triple Tower Architecture Perception tower Model tower Simple reflexes Action tower Sensory signals Motor commands

6 Sloman s CogAff architecture Meta-management (reflective processes) Deliberative reasoning ( what if mechanisms) Reactive mechanisms Attitudes Standards Motives LTM Moods Filters perception central processing extra mechanisms 219 6