Applying Anticipatory Networks to Scenario Planning and Backcasting in Technological Foresight

Transcription

1 Applying Anticipatory Networks to Scenario Planning and Backcasting in Technological Foresight Andrzej M.J. Skulimowski Decision Sciences Department, AGH University of Science & Technology, Kraków, Poland International Centre for Decision Sciences and Forecasting Progress & Business Foundation 5th FTA Conference, Brussels November 27-28, 2014

2 Organisation of this talk: 1. Motivations: complex decision-making problems for middle- and long-term strategic planning 2. Future exploration methodology: forecasting, foresight, scenarios, planning, backcasting, roadmapping 3. Technological evolution as a modelling background 4. An introduction to modelling the future with anticipatory networks 5. Network algorithms for reverse planning 6. Anticipatory networks as superanticipatory systems 7. Hybrid anticipatory networks 8. Anticipatory modelling process: from scenario planning to roadmapping 9. Case study: Anticipatory planning of operations of a Regional Creativity Support Centre 10. Final Conclusions and Remarks

3 Motivations: Practice-oriented: o Strategic planning: a need to fuse and implement the quantitative results of foresight, normative scenarios, and impact analysis o Better exploration of foresight outcomes o Analysis of sustainability problems: multicriteria decisions for better future, environmental models for sustainability assessment o Filtering ICT/AI development scenarios only those nondominated remain Theory-oriented: o Vector-valued utilities in group decision-making o Multi-step Stackelberg games, multicriteria n-level programming o Discrete event systems control in causal networks with feedback

4 Multicriteria Optimisers A scheme of operation of a simple multicriteria optimiser O corresponding to the problem (F:U E) min(p) with the solution set Π(U,F,P) (MO) F U P OPTIMIZER O=X(F,U,P) X(F,U,P) F(X(F,U,P) O may be endowed additionally with an interactive selection mechanism ψ: Π(U,F,P) x I Π(U,F,P), where Π(U,F,P) is the set of nondominated points for the problem (MO) and I is the set of external states of knowledge. ψ models the selection of a unique compromise solution from the set of nondominated solutions

5 Networks of Optimizers: Simple Chains Linked multicriteria optimizers: O 1 φ(1) O 2 φ(2) O 3 φ(3) φ(n-1) O n (2), Causality model: a) linking functions φ force the choice of a solution in subsequent problems b) linking functions φ influence the sets U i only c) φ can also influence the choice of criteria F i and the selection rules ψ

6 N-level model of consequences The decision process: u i (u i-1 ) for i=1,..n, :=Y i F i-1

7 Anticipatory decision problems in networks of optimisers Assumptions: 1. Solution to an optimisation problem O p may influence M p independent subsequent problems 2. An optimisation problem O r may be influenced by N r preceding problems 3. The influence aggregation rules can be defined for each influenced optimiser (e.g. as intersection of the sets of feasible alternatives, each one imposed by a different preceding optimiser) Explorative Scenarios and Forecasts can be used simultaneously to support decisions, when the decisions concern a middle- or long-term future. Scenarios can be external-events-driven, when included in solution models, they allow to generate decision rules

8 Problem formulation for the networks anticipatory feedback with the future Definition 1. Let O i i=0,1,,n, be a chain of optimizers with causal relations given by the restrictions of the scope of admissible decisions in U i+1 such that if u i is an admissible decision for the optimizer O i then u i+1 is an admissible decision in U i+1 iff u i+1 (i)(u i ), where φ(i):=y(i) F i-1 is a multifunction describing the restrictions on admissible choice of solutions in optimization problems subsequent to O i. Moreover, let us assume that all elements of U 0 are admissible. Any sequence of admissible solutions (u 0,m(0), u 1,m(1),, u N,m(N) ) will be called an admissible chain. Definition 2. For a chain of optimizers O i i=0,1,,n the anticipatory feedback condition (FC) is defined as the requirement that for j J:= {J(i):i=0,,N-1} and certain family of subsets {V i,j } j J. the feedback at O i is realized so that the decision-maker at O i strives to select the solution which guarantees that the the solutions selected at a future decision node O j belong to {V i,j } j J for all j J(i).

9 Scenario Filtering Problems Problem 1. Find the set of all admissible chains that fulfil additionally the anticipatory feedback condition FC (i.e. anticipatory scenarios). A problem with the relaxed feedback condition: Problem 2. Find the set of all admissible chains that maximize the following function: g(u 0,,u N ):= i J(0) h(u i, q(i))w 0,i (RFC1) such that for all i, 1 i<n, the truncated admissible chain (u i,,u N ) maximizes the function g(u i,,u N ):= j J(i) h(u j, q(j))w i,j where h is a quantitative measure of satisfaction of FC, e.g. h(u i,q(i)):= F i-1 (u i )-q(i), and w i,j are positive relevance coefficients for the feedback relation between the optimizers O i and O j. (RFC2) Problem 3. Find the set of all admissible chains that fulfill the feedback condition or the conditions RFC1,2 and start at a nondominated point at U 0

10 Anticipatory Networks and Causality Definition 3. A scenario which is a solution to Problem 3 will be called an anticipatory scenario Definition 4. Any two optimizers O m =X 1 (U,F,P) and O n =X 2 (W,G,R) are in the causal influence relation if there exist two different outputs from O m, x 1,x 2 X 1, such that either the choice in O n is restricted to two different subsets of W that depend on choice of x 1 or x 2 in O m, or if the solution selection rule R or the criterion G are modified in different manner depending on the choice of x 1 or x 2. Definition 5. An anticipatory network (of optimizers) is a causal network (i.e. network with at least two optimizers in the causal influence relation with at least one additional anticipatory feedback relation.

11 A computational example (1): a chain of five multicriteria optimization problems (U i,f i )

12 A computational example (2) The multifuctions (i) are defined explicitly by showing allowed choices

13 Networks of Optimizers Anticipatory Trees A tree of optimizers with future feedback consisting of 10 elements O i =(U i,f i,p i ), i=0,1, 9, and four chains, with F i :=id Ui, causal relations defined by multifunctions Y i, and seven essential future feedback relations (solid arrows). The dotted arrow between U 3 and U 9 is an irrelevant anticipatory feedback (no causal relation).

14 Anticipatory Networks in the general case

15 Hybrid Anticipatory Networks new units The basic decision units that may occur in a hybrid anticipatory network: simple box (a) - multiple-input optimizer with output possibly influenced by the states of nature N, triangle (b) - the pre-determined algorithmic decision unit, where the decision may additionally depend on the states of nature N, rounded box (c) random decision is selected based on known inputs and an output distribution function, subdivided box (d) 2-player non-cooperative game unit

16 A Hybrid Anticipatory Network

17 Further applications: Backcasting and sustainability The common problem: how to reach a desired state in the future Backcasting Anticipatory feedback models Source:

18 Standard backcasting vs. anticipatory networks Specification A typical backcasting procedure Anticipatory modelling Common goals To assess the feasibility of given visions of the future To support strategic planning projects that involve multiple strategies Time scale Core target analysis Strategic planning Interaction Determine the reverse planning horizon based on information provided by experts and stakeholders Define future ideal states (normative scenarios), and feasible actions that may be applied to reach them Engage experts and stakeholders to take part in the action planning. Use panels, workshops, brainstorming, etc. to elicit their opinions All steps in the procedure should be repeated with updated information and forecasts until the preference thresholds set by stakeholders and decision makers are met In addition to expert information: AN allows the modeller to determine the maximum grade of the network and derive the anticipation horizon out of it Define the criteria and preference structures as well as reference points for relaxing anticipatory feedback conditions Use AN algorithms to calculate compromise solutions along anticipatory chains. The information provided while building the AN generates the solution Once built, the AN remains stable during the analysis. The interaction touches upon the presentation of different nondominated anticipatory chains until the present-time decision makers are satisfied

19 Case Study: Anticipatory planning for a Regional Creativity Support Centre The case studied within the foresight project SCETIST,

20 Case Study: Anticipatory planning for a Regional Creativity Support Centre Stage II: Long-term planning The case studied within the foresight project SCETIST,

21 References: R. Rosen (1985). Anticipatory Systems - Philosophical, Mathematical and Methodological Foundations, Pergamon Press, London (2nd Edition, Springer, 2012). A.M.J. Skulimowski (1985). Solving Vector Optimization Problems via Multilevel Analysis of Foreseen Consequences. Found. Control Engrg., 10, No. 1, (available from A.M.J. Skulimowski (2012). Hybrid Anticipatory Networks. 11th ICAISC, Zakopane 2012, LNAI 7268, Springer, pp , A.M.J. Skulimowski (2013). Exploring the future with anticipatory networks, AIP Conf. Proc. 1510, pp ; doi: A.M.J. Skulimowski (2014). Anticipatory Network Models of Multicriteria Decision-Making Processes. Int.J.Systems Sci, 45(1), 39-59,

22 THANK YOU FOR YOUR ATTENTION! INQUIRIES: ams (at) agh.edu.pl, This research has been supported within the foresight project No. WND- POIG /09: Scenarios and development trends of selected information society technologies until 2025 (SCETIST) funded by the ERDF within the Innovative Economy Operational Programme,