Mul$agent Planning for Social Computa$on

Transcription

1 Mul$agent Planning for Social Computa$on Michael Rovatsos School of Informa$cs, University of Edinburgh CETINIA Invited Lecture Universidad Rey Juan Carlos, Madrid 10 th October 2013

2 Social Computa$on! Large- scale man- machine collec$ves that perform complex computa$ons mediated by digital media! OIen based on crowdsourcing and/or human computa$on, e.g. Captcha, Galaxy Zoo, TaskRabbit, Kickstarter! Offers a new vision for collec$vely intelligent systems, where people and ar$ficial agents complement each other! Surprisingly, mul$agent systems technologies are not used much for the engineering of these systems! Conversely, exis$ng systems have been built in ad hoc fashion, achieving good design rather a black art 2

3 The SmartSociety Project! 4- year 6.8M EU FP7 FET Integrated Project, co- ordinated by University of Trento! Aim: building hybrid and diversity- aware collec$ve adap$ve systems to solve challenging societal problems! Edinburgh focus within the project: social orchestra9on & composi9onality of social computa$on systems! lightweight architectures for composing workflows on the Web! enabling composi$onality with human and machine help! iden$fying emergent structure and evolving system design

4 The relevance of planning! Planning is a key ability of intelligent systems! how do I achieve a goal given what s true now and what parts of the world I can influence?! algorithmically: the problem of genera$ng ac$on sequences that will reach a state with certain proper$es! Relevance for social computa$on: useful to! describe! synthesise! verify! execute complex distributed tasks 4

5 Structure of this talk! Part I: Introduc$on! social computa$on, and why planning is relevant! Part II: Mul$agent planning problems & solu$ons! several planning- related problems in mul$agent sefngs! present them as canonical within a uniform frameworks! Part III: Applica$on to social computa$on problems! poten$al uses of each proposed method! challenges and opportuni$es! Part IV: Conclusions I will use transporta/on domains as examples throughout (e.g route planning, portage task planning, ridesharing) 5

6 Single- Agent Planning! For single- agent planning, there exist simple and general formula$ons of planning problems! STRIPS- based classical planning problem P=<F,I,A,G> with fluents F, ini$al state I, ac$ons A and goal G! Fluents are proposi$onal proper$es of states, states are sets of these! Ac$ons have the form <a,pre,eff> where pre F and eff=(add,del)! Fluents add F and del F are added to/deleted from the current state when a is executed 6

7 Single- Agent Planning! State transi$on func$on yields S =S \ del(a) add(a)when a is executed in S! Blocks World example:! ac$on a=stack(x,y)! pre(a)={clear(y), Holding(X)}! add(a)={on(x,y), ArmEmpty}! del(a)={clear(y), Holding(X)}! A plan p=<a 1,...,a n > is a solu$on to planning problem P if execu$on of a 1,...,a n from I yields a state S and G S! Various extensions to this: conformant/condi$onal planning, concurrency and scheduling, temporal planning, planning with preferences and side condi$ons 7

8 Single- Agent Planning! Planning made a lot of progress because of! common problem formula$on! scalable algorithms! benchmarks for empirical evalua$on! In mul$agent planning, addi$onal complica$ons:! concurrent ac$on, different views, different goals! coordina$ng the planning ac$vity itself! This makes the problem much much harder... 8

9 Mul$agent planning! Differently from single- agent planning, highly fragmented! Problems addressed include:! centralised planning with concurrent ac$ons! plan merging from individual agent plans! planning- $me co- ordina$on of planning agents! centralised/decentralised strategic planning! con$nuous planning, execu$on, and co- ordina$on 9

10 Our work in mul$agent planning! We try to focus on mul$agent planning problems that address different issues which we consider canonical! These can be defined using just simple varia$ons of the standard formula$on of the classical planning problem! I will discuss work on four problems:! mul$- perspec$ve planning! mul$- objec$ve planning! automated norm synthesis! automated agent decomposi$on 10

11 Mul$- Perspec$ve Planning! Agents disagree about ini$al state and ac$on defini$ons, but share goal: P i =<F,A i,i i,g>! Acceptable planning problem: plan p is acceptable wrt KB 1 and KB 2 iff KB 1 = p and KB 2 = p! Belesio$s & MR developed argumenta$on- based method based on evalua$ng individual agents proposals, solu$on = a plan that can be defended! Scalability achieved by using off- the- shelf single- agent planner for sub- tasks in the argumenta$on process 11

12 Mul$- Perspec$ve Planning Assign Roles Re-evaluate Arguments FALSE Identify Relevant Conflicts FALSE Proposal Proposal Constructed Validate Validation Succeeds FALSE TRUE Argumentation Is Defendable TRUE TRUE Resolve Conflicts Plan proposal generated by single agent (with any planner) Dispute in case of disagreement, argumenta$on follows Ends in successful defence of ini$al proposal or rejec$on Middle- ground between classical and conformant planning 12

13 Example! Robot gridworld domain PRO 2 1 o r t OPP t r o 1 2 pickup drop (1): P (, pickup,, drop) s 0 s 1 s 2 s 3 s 4 (4): HI ( have(r, o)) (6): HI ( at(o, 1, 2)) (5): DA (pickup, at(o, 1, 2),s 1 ) (3): HE (pickup, hold(r, o),s 3 ) (7): Belief (2): GNA (at(o, 2, 2)) proponent moves opponent moves 13

14 Applica$on: ArguDem! A demonstrator for helping robots navigate: 14

15 Mul$- Objec$ve Planning! Agents have independent goals: P=<F,A,I,G i >! Strategic problem, acceptability based on no$ons of stability and equilibrium! Problem depends on whether contracts can be enforced and u$lity can be transferred! Like concurrent planning with addi$onal constraints on plan cost to individuals! Hard to define meaningful solu$on concepts if goals incompa$ble or agents untrustworthy 15

16 Example Delivery domain Note: focus on domains where enumera$on of plans is impossible 16

17 Planning for Self- Interested Agents! Best- Response Planning (Jonsson & MR):! itera$ve method of op$mising agents individual plans without breaking others plans! computes equilibrium plans fast in conges$on games, restricted to interac$ons regarding cost! useful for plan op$misa$on in unrestricted domains! Later also extended to include compress- and- expand algorithm to produce ini$al concurrent plan! only for domains where agents can achieve their individual goals alone; but useful for plan cost op$misa$on 17

18 Applica$on: Ridesharing! Hrncir & Hlavaty s system uses BRP to calculate travel routes using real- world UK public transporta$on data and private cars (>200K connec$ons) 18

19 Automated norm synthesis! Avoiding undesirable states in a system regardless of agents planning ac$vi$es! Automated synthesis = given a planning domain, calculate a set of prohibi$ons for agents! Automated synthesis of such norms is NP- hard in enumerated state systems! Exis$ng methods don t exploit abstrac$ons of proposi$onal/first- order domain theories 19

20 Automated norm synthesis! Christelis & MR developed CRS algorithm based on forward- backward search around conflict states guaranteeing full goal accessibility! Pruning techniques and use of single- agent performant planners result in highly scalable methods! Our method: find detours around conflict states by local search in generalised state spaces 20

21 Automated norm synthesis! Iterated process of forward- backward search around conflict state specifica$on: Not beser than full state- space search in the worst case but we can oien get lucky 21

22 Example! Tunnel world example: Tunnel 1 Tunnel 2 Tunnel! Agents entering tunnels have to leave them out the opposite end immediately (on entering tunnel, future crash not avoidable)! Our algorithm solves this by compu$ng a general norm if you are next to a tunnel and another agent is at the opposite end, don t enter the tunnel 22

23 Automated Agent Decomposi$on! Everything described so far assumes explicit representa$on of agents in problem formula$on! But what makes a domain mul$agent, if all you knew was its classical specifica$on P=<F,I,A,G>?! Many benchmark problems for (single- agent) planning problems are mul$agent to the human eye but does that even maser for an algorithm?! Yes, because! we can build plan search algorithms that exploit this structure! it provides an elegant defini$on of agents 23

24 Automated Agent Decomposi$on! Crosby developed automated decomposi$on technique that iden$fies meaningful agent variable decomposi$ons 24

25 The ADP Algorithm! Automated agent decomposi$on algorithm! Planning algorithm that exploits this structure! Iden$fies decomposable domains fast and reliably! ADP algorithm outperforms state- of- the- art planners on those domains! Relevant for about 40% of common benchmark (IPC) domains 25

26 Back to social Computa$on! All these methods can be used in social computa$on, e.g. ridesharing scenario! Mul$- perspec$ve planning: resolve disagreements among par$cipants in a structured, verifiable way! different travellers have different knowledge about feasibility and quality of trips, e.g. punctuality, reliability, comfort! oien there is conflic$ng knowledge from different informa$on sources, or preferences of individuals are unknown a priori! incomplete informa$on can be refined by humans who incrementally curate the global travel database! challenge: argumenta$on complexity grows exponen$ally with number of possible different views 26

27 Back to social computa$on! Mul$- objec$ve planning: compute travel routes that are acceptable for all travellers involved! reduces probability of individuals devia$ng from joint plan (though collusion among sub- groups is s$ll possible)! allows for broad range of cost func$ons and taking account of external extensives (e.g. to influence traveller behaviour or avoid conges$on)! may help convince people of benefits of ridesharing by helping to navigate unmanageable space of possible rides! challenge: elici$ng preferences in a lightweight yet reliable way, execu$on monitoring despite par$al observability 27

28 Back to social computa$on! Automated norm synthesis: ensure global constraints (especially regarding safety) are met despite flexible design! where full reachability of states can be maintained, norms can be proven not to constrain individual travellers mobility! poten$al addi$onal cost incurred can be es$mated by comparing plans under norm ac$va$on and deac$va$on! important for reassuring travellers in a sensi$ve domain (passenger safety & predictability of trip and its cost essen$al)! challenge: nego$a$ng tradeoffs with user communi$es where full reachability is lost due to synthesised norms 28

29 Back to social computa$on! Automated agent decomposi$on: useful to iden$fy hidden locality and parallelism in a domain! most relevant for complex combinatorial tasks, e.g. planning a set of overlapping rideshares for a larger popula$on! can be used to deal with side- effects and undiscovered conflicts and scale up plan computa$on to larger collec$ves! challenge: finding the right abstrac$ons for group ac$on that will reveal and resolve poten$al cross- group conflicts 29

30 Back to social computa$on! Most challenges when applying these techniques to social computa$on are to do with! Large- scale agent popula$ons interac$ng sporadically! Refinement through con$nually elicited human knowledge! The unpredictable nature of human users! Desiderata for future research:! Approaches must be able to deal with collec$ves! Algorithms must allow for varying levels of detail! Representa$ons must be adapted to past experience! If methods don t work for humans, forget about them! 30

31 ESSENCE - Evolu$on of Shared Seman$cs in Computa$onal Environments! 4- year, 4M Marie Curie Ini$al Training Network, co- ordinated by Edinburgh (starts 1 st November)! Aim: to exploit human methods for nego$a$ng, sharing, and evolving meaning in computa$onal systems! High- profile consor$um, from pure logic and game theory to natural language processing and sensor networks! Focus in Edinburgh: Communica$on planning from heterogeneous sensor data and ontology learning! By the way, we have funding for 11 PhD students and 4 post- docs ($$$$$ - but you have to go abroad!)! Visit network.eu soon, or contact me now (mrovatso@inf.ed.ac.uk)

32 Conclusions! Argued for social computa$on as an exci$ng and challenging domain for mul$agent planning! Massive poten$al impact, though currently lisle use of mul$agent systems techniques! Gave examples from previous work in our group that may be useful star$ng points! Raises challenges which can only be addressed if we adapt certain methodological assump$ons - think out of the box! The future is bright, but not easy to conquer! 32

33 Thanks for your asen$on!! Ques$ons? 33