To design a rational agent you must first specify its TASK ENVIRONMENT. PEAS is the standard four-part description: P — PERFORMANCE MEASURE The criteria by which success is judged. Recall from the agents topic that this is supplied by the DESIGNER and that specifying it badly produces perverse behaviour. It is the first item for a reason: everything else is chosen to serve it. E — ENVIRONMENT What the agent operates in — the world it perceives and affects, including other agents. A — ACTUATORS The means by which the agent acts on the environment. This defines the ACTION SPACE, and constraining it is the most reliable way to bound what an agent can do. S — SENSORS The means by which the agent perceives. This defines the PERCEPT SPACE, and it determines what the agent can possibly know — no amount of cleverness compensates for a missing sensor. THE ORDER IS DELIBERATE. Performance measure first, because it determines what counts as a relevant percept and a useful action. Designers who start from "what sensors do we have?" build systems that measure what is easy rather than what matters. THE CANONICAL EXAMPLE — an automated taxi driver: P safe, fast, legal, comfortable trip; maximise profits; minimise impact on other road users E roads, other traffic, police, pedestrians, customers, weather A steering, accelerator, brake, signal, horn, display, speech S cameras, radar, lidar, speedometer, GPS, engine and fuel sensors, accelerometer, microphone, keyboard NOTE THE PERFORMANCE MEASURE'S INTERNAL CONFLICT: safe and fast oppose each other, as do profit and minimising impact on others. THE MEASURE MUST ENCODE THE TRADE-OFF, and this is the same conflicting-non-functional-requirements problem the architecture topic described. A measure listing conflicting goals without weights has not finished the job. MORE PEAS DESCRIPTIONS, worth knowing as exam material: MEDICAL DIAGNOSIS SYSTEM P healthy patient, reduced costs, minimised lawsuits E patient, hospital, staff A display of questions, tests, diagnoses, treatments, referrals S keyboard entry of symptoms, findings, patient's answers SATELLITE IMAGE ANALYSIS P correct categorisation of terrain E orbiting satellite, downlink, weather A display of scene categorisation S colour pixel arrays PART-PICKING ROBOT P percentage of parts in the correct bins E conveyor belt with parts, bins A jointed arm and hand S camera, joint angle sensors REFINERY CONTROLLER P purity, yield, safety E refinery, raw materials, operators A valves, pumps, heaters, stirrers, displays S temperature, pressure, flow, chemical sensors INTERACTIVE ENGLISH TUTOR P student's score on a test E set of students, testing agency A display of exercises, corrections, suggestions S keyboard entry, student's answers LOOK AT THE ENGLISH TUTOR'S PERFORMANCE MEASURE and apply the warning from the agents topic: "student's score on a test" is gameable. An agent optimising it will teach to the test, drill question formats rather than English, and may discover that easier tests raise scores. THE HONEST MEASURE — actual language competence — is harder to observe, which is exactly why the gameable proxy gets chosen.

SPECIFY an agent that recommends which topic a student should study next — the running example from the agents topic, now done properly. ── P: PERFORMANCE MEASURE ─────────────────────────────────── THE REAL GOAL: the student passes the exam with the highest achievable mark. MEASURABLE FORM: expected exam mark, approximated by Σ over topics [ exam_weight(topic) × P(recalled at exam date | study history) ] minus a penalty for study time, since the student's time is finite. REJECTED CANDIDATES, and why — this is the valuable part: ✗ "topics ticked" → rewards short easy topics ✗ "recommendations followed" → rewards recommending what is easiest to accept ✗ "hours studied" → rewards keeping the student working, not learning ✗ "articles read to 100%" → rewards scrolling, which is a measure the agent can influence without teaching anything EACH REJECTED MEASURE IS A MEASURE THE AGENT COULD SATISFY WITHOUT DOING THE JOB. Finding four of them in one afternoon is normal, and is the argument for doing this exercise. ── E: ENVIRONMENT ─────────────────────────────────────────── the student (whose knowledge state is HIDDEN and can only be inferred) · the syllabus of 519 topics with exam weights · the exam date · the calendar (available study time) · the student's device PROPERTIES (from the environment types topic): partially observable — the agent sees ticks and reading depth, never comprehension stochastic — the student may or may not study, and may forget at an individual rate sequential — today's recommendation changes tomorrow's state dynamic — the exam approaches whether the agent acts or not single-agent · discrete · known rules ── A: ACTUATORS ───────────────────────────────────────────── display a recommended topic with a reason · display a revision prompt · display "nothing due — rest" · reorder the subject list · send a notification DELIBERATELY EXCLUDED, and the exclusions are the design: ✗ marking a topic complete on the student's behalf — that is the student's claim about their own knowledge, and an agent asserting it corrupts the very data the agent learns from ✗ changing the exam date ✗ hiding topics from the syllabus CONSTRAINING THE ACTION SPACE IS THE AUTONOMY/PREDICTABILITY MITIGATION from the properties topic, applied concretely: an agent that cannot mark topics complete will never find a clever reason to. ── S: SENSORS ─────────────────────────────────────────────── which topics are ticked and when · reading depth and time per article (the ReadingProgress data) · notes written · the exam date · which recommendations were shown and whether they were acted on WHAT THE AGENT CANNOT SENSE, and this is the limiting fact: whether the student UNDERSTOOD anything. Ticking a topic and reading it to the end are proxies for comprehension, and a student can do both while learning nothing. THE DESIGN CONSEQUENCE: because comprehension is unobservable, the agent's model is necessarily uncertain, which is why the performance measure is stated as a PROBABILITY of recall rather than a fact. A SENSOR THAT DOES NOT EXIST CANNOT BE COMPENSATED FOR BY A BETTER ALGORITHM — the honest response is to add a sensor (a self-test, a quick quiz) rather than to infer harder. WHAT THE EXERCISE PRODUCED: one measurable objective, with four rejected alternatives on record six environment properties, which determine the agent structure needed five actuators and three explicit exclusions with reasons five sensors and one named blind spot with its consequence NONE OF IT REQUIRED CODE, and every item would otherwise have been decided implicitly. THAT IS WHAT PEAS IS FOR.

THE COMMON MISTAKES, each with the correction: STARTING WITH SENSORS INSTEAD OF THE MEASURE "We have this data, what can we do with it?" produces a system that optimises whatever the available data measures. Start from what success means, then ask what you would need to sense in order to tell. A PERFORMANCE MEASURE THE AGENT CAN INFLUENCE DIRECTLY Any measure the agent can move without doing the work will be moved that way. Measure the ENVIRONMENT'S state, not the agent's report of it — Goodhart's law, again. CONFUSING ACTUATORS WITH CAPABILITIES "Actuators: machine learning" is not an actuator. An actuator is a means of ACTING ON THE ENVIRONMENT — a display, a motor, a message sent, a record written. LEAVING THE ENVIRONMENT VAGUE "The environment: the internet" tells you nothing. The useful description names what the agent perceives and affects, and includes the OTHER AGENTS — because their presence changes the problem from optimisation to game theory. OMITTING THE ADVERSARIAL PART OF THE ENVIRONMENT A spam filter's environment includes SPAMMERS WHO ADAPT. A fraud detector's includes fraudsters who probe it. Modelling these as static noise rather than as responsive agents is why such systems degrade: the environment is not merely dynamic, it is dynamic BECAUSE OF THE AGENT'S OWN ACTIONS. NOT WRITING DOWN THE EXCLUSIONS The actions an agent may NOT take are often the most important part of the specification, and they are the part that never gets recorded unless the template asks for them. PEAS AND THE REST OF THE DESIGN — how the four items propagate: P determines whether a goal-based agent suffices or a UTILITY function is needed. A measure with a single binary criterion permits a goal-based agent; a measure trading off several dimensions requires utilities. E determines the required agent STRUCTURE. Partially observable → the agent needs internal state, so a simple reflex agent is ruled out. Stochastic → it must handle uncertainty. Multi-agent → it may need to model others. A bounds the SEARCH SPACE — the branching factor of every search algorithm in ACtE0902 is the number of applicable actions. S bounds what is KNOWABLE, and therefore what the agent's state estimate can possibly contain. SO PEAS IS NOT A FORM TO FILL IN AND FILE. Each item constrains the design that follows, which is why it comes first — the same relationship the requirements have to architecture in the software engineering unit.