index.md (2044B)
1 +++ 2 title = "Rational agents" 3 template = 'page-math.html' 4 +++ 5 # Rational agents 6 7 "A rational agent chooses whichever action maximizes the expected value of the performance measure given the percept sequence to date and prior environment knowledge." 8 9 ## Agents 10 agent function maps percept sequence to actions ($f: P* \rightarrow A$) 11 12 function is internally represented by agent program 13 14 program runs on physical architecture to produce f 15 16 ## Rationality 17 what is rational at a specific time depends on: 18 * expected value of performance measure -- heuristics 19 * actions and choices -- search 20 * percept sequence to date -- learning 21 * prior environment-- KR 22 23 rationality is not omniscience or perfection 24 25 ## Task environments 26 27 to design rational agent, we must specify environment (PEAS): 28 * performance: safety, destination, profits, legality, comfort 29 * environment: streets, traffic, pedestrians, weather 30 * actuators: steering, accelerating, brake, horn, speaker/display 31 * sensors: video, sonar, speedometer, etc. 32 33 environment types: 34 * observable: fully (can detect all relevant aspects with sensors) or partially 35 * deterministic: (yes or no) 36 * static: (yes, no, semi) 37 * discrete: (yes or no) 38 * single-agent: (yes or no) 39 40  41 42 For Schnapsen: 43 * observable: not fully 44 * deterministic: yes 45 * static: yes 46 * discrete: yes 47 * single-agent: no 48 49 ## Agent types 50 51 ### Simple Reflex 52 select action on basis of _only the current percept_ 53 54 large reduction in possible percept/action situations 55 56 implemented using condition-action rules 57 58 only works if environment is fully observable, otherwise may result in infinite loops. 59 60 ### Reflex & State 61 to tackle partially observable environments, maintain internal state 62 63 over time, update state using world knowledge. 64 65 ### Goal-Based 66 agent needs a goal to know the desirable situations 67 68 future is taken into account 69 70 ### Learning 71 teach agents instead of instructing them 72 73 very robust toward initially unknown environments. 74