class: center, middle # Artificial Intelligence: Planning ### Plan Recognition --- <img src="/CI-2807/assets/img/posterII2019.png" width="100%"/> --- # Review: The Planning Problem A planning problem consists of three parts: * A definition of the current state of the world * A definition of a desired state of the world * A definition of the actions the agent can take --- # Observation * Imagine we want to cooperate with an agent (or a human) * We can see which actions they are performing * Now we need to figure out what they want to achieve --- class: medium # Plan Recognition * Goal Recognition: Given some observed actions, figure out which goal the observed agent is trying to achieve * Plan Recognition: Given some observed actions, figure out which plan the observed agent is likely to follow * Activity Recognition: Given some observed actions, figure out which of them are relevant to an activity performed by the observed agent Confusing: All of them were at some point or another called "Plan Recognition" --- # Intention Recognition * More confusion: Some people also use "intention recognition" * Unfortunately, this has some non-research connotations of "mind reading" * Philosophers also have their own (different) ideas of what "intentions" are * We will be clear, and us "Plan Recognition" for the task of determining the *plan* that the observed agent is (likely) following --- # Applications * Games (cooperative and competitive) * Assistive care for the elderly * Software help * Security * Cooking * Etc. --- # Plan Recognition Approaches * Plan library: Match observed actions against known plans * Statistical methods: Collect data from many users to predict likely follower actions * Planning: What plan would *you* execute if you had done the observed actions --- # Plan Recognition: Example <img src="/PF-3335/assets/img/planrecognition.png" width="80%"/> --- class: center, middle # Plan Recognition as Planning --- # Assumptions * We know the initial state of the world * We observe complete actions * We observe them in order * We know what goals the observed agent could potentially be trying to achieve * We believe the observed agent follows the optimal plan --- # The Plan Recognition Problem Given: * A set of observed actions * A set of potential goals Find: * A set of plans the agent is likely to follow --- # A different perspective * If we know the initial state and the potential goals, we can just find a plan for each of them * Then we compare the plans we found with the actions the agent performed * If there's a match, we found a potential solution * What could possibly go wrong? --- # Optimality * We assume the agent follows "the" optimal plan * What if there is more than one option? * The agent may find one optimal plan * We find a different one that is also optimal * Those two don't match ... --- class: medium # Incorporating Observed Actions * What we want is that our plans *have to* include the observed actions * We can include this in the goals! * We create a new predicate `observed`, which has one parameter: an action/step number * For each observed action `i` we create a **new** action that has an additional precondition `(not (observed i))` and an additional effect `(observed i)` * We add **all** `(observed i)` atoms to each of the goals --- # Plan Recognition as Planning * To perform the plan recognition, we now have to run the planner **twice** for each goal * One time, we try to find a plan given the initial state, with the unmodified actions and goal * The second time, we try to find a plan with the modified actions and goal * If (and only if) these two plans have the same length, do we consider the resulting plan a possibility --- # Performance * As mentioned, we need to use an optimal planner *twice* for each possible goal * This will be computationally expensive * How about we use a heuristic planner that may not be optimal? * We may overlook a few possible plans, or find a few extra ones --- # Evaluation * Let's say we have produced a set of plans that we believe the agent is following * How do we determine if we "did well"? * What was our goal? * What can we expect? --- # Evaluation: Precision and Recall * Precision: What percentage of produced plans are actually reasonable? * Recall: What percentage of reasonable plans do we produce? Both of these measures are important! We can use them to find out how good our "non-optimal" planner did --- # Plan Recognition: Ideal Solution? * Precision and Recall tell us how well an algorithm does * But what do we actually want? * Also, does the *domain* have any effect on how good we can do? * What if it's impossible to predict what the observed agent is trying to do? --- # Worst Case Distinctiveness <img src="/PF-3335/assets/img/wcd.png" width="60%"/> --- # Worst Case Distinctiveness * Worse Case Distinctiveness is a measure of a *domain* * It basically says how long we have to wait *in the worst case* before we can figure out what the observed agent is up to * This can be arbitrarily "bad" (imagine an "infinite" corridor with two doors at the end ...) * In practice we may want to modify our domain to improve the situation --- # Worst Case Distinctiveness: Domain Modification <img src="/PF-3335/assets/img/wcdfix.png" width="100%"/> --- class: center, middle # Extensions --- # Extensions * Partial plans * Observations * Noise * Continual Plan Recognition * Suboptimality --- # Partial Plans <img src="/PF-3335/assets/img/planrecognition_partial.png" width="80%"/> --- # Observations * So far we have assumed we can observe actions * What if we can just observe the world? * We will have to deduce which actions the agent performed in order to enact the changes we observe * There may be multiple options for some changes --- # Noise * Maybe we can't observe the *full* action, but only parts of it * Or we can't observe the *full* world * We may have to use some sort of fuzzy reasoning (or a possible worlds model!) to represent the world --- # Continual Plan Recognition * In many domains, like games, we will perform plan recognition over and over (e.g. every turn) * But many of the computations will be repeated * If we design a planner that keeps track of expanded states (and discards the ones corresponding to actions that did not happen), we can make these computations more efficient --- # Homework * No more homework * Next week we will wrap up the class with some applications and an outlook * There will also be a discussion of the planning competition --- class: small # References * [Plan Activity and Intent Recognition Tutorial](http://www.planrec.org/Tutorial/Resources_files/pair-tutorial.pdf) * [Resources for Plan, Activity, Behavior, and Intent Recognition Research](http://www.planrec.org/Resources.html) * [Plan Recognition as Planning](https://sites.google.com/site/prasplanning/home) * [Goal Recognition Design](https://www.aaai.org/ocs/index.php/ICAPS/ICAPS14/paper/viewFile/7814/8023)