A Case Study Exploring Suitability of Bottom Up Modelling and Actor-based Simulation for Decision Making Souvik Barat Asha Rajbhoj Prashant Kumar Tata Consultancy Services Research Tata Consultancy Services Research Tata Consultancy Services Research Pune, India Pune, India Pune, India souvik.barat@tcs.com asha.rajbhoj@tcs.com kumar.prashant10@tcs.com Vinay Kulkarni Tata Consultancy Services Research Pune, India vinay.vkulkarni@tcs.com ABSTRACT individual elements and their interactions in precise form as oppose to the overall system behaviour. Conceptually, the bottom-up Traditionally, the top-down design method and analysis techniques, approach relies on emergentism [4] as advocated in actor model of such as system dynamic model, have been used extensively for computation [5, 6], and agent-based systems [7]. understanding complex systems. In top-down approach, a system is Traditionally, the top-down approach is popular choice (as compare specified in terms of global state and the desired analyses are to the bottom-up approach) for analysing and understanding the performed using aggregated macro-behaviour that represents the complex systems in the context of critical business needs such as overall system. Essentially, the individual elements and their decision making activities. Existing modelling and analysis tools peculiarities are not differentiated with an assumption that the supporting top-down approach are extremely efficient for inherent dynamics of the overall system is precisely known to the describing and simulating the aggregated system behaviour. system modellers. This paper, in contrast, presents a case wherein However, they are not appropriate for precise understanding of the system behaviour emerges from the individual elements and complex and dynamic system that exhibits emergent behaviour and their interactions. The paper further demonstrates the usability of deals with large number of socio-technical [8] elements having bottom up approach, actor based modelling abstraction, and actor adaptive, autonomous and dynamic behaviours. based simulation technique to understand complex systems (with In this paper, we demonstrate the effectiveness of bottom-up emergent behaviour) using a case study on decision making of a modelling approach to understand a complex and dynamic system Research and Innovation (R&I) organisation. (with emergent behaviour) using a case study that illustrates decision making of an industrial Research and Innovation (R&I) KEYWORDS organisation. The goals of R&I organisation is to convert Complex dynamic decision making; Enterprise modeling; Bottom- innovative ideas into business offerings, and make significant up modelling; Simulation; Actor-based simulation. scholastic impacts to the research community (through publications and patents). In this context, the behaviour of the overall R&I 1 INTRODUCTION organisation is not well-defined - rather it emerges from the activities of the individual researchers. Moreover, the progress of Two design alternatives, named top-down approach and bottom-up the organisation largely relies on the effective utilization of the approach, exist for specifying and analysing complex systems [1]. researchers within the dynamic compositions structure where they In top-down approach, a system is visualised in terms of global operate (i.e., research project). state and the behaviour is represented using aggregated macro- We adopt the concept of actor model of computation [6] to behaviour of the system elements. For example, the System represent constituent elements, such as the research projects and Dynamics (SD) model [2] uses the concepts of stocks, flows and researchers, of R&I organisation; formulate simulation setting by information to represent system state and system level nonlinearity, allowing these elements to interact with each other (as oppose to feedback loops and the time delays. The behaviour is described describing overall R&I organisation specification); and use actor using differential equations. In principal, these modeling elements based simulation technique to observe the emergent behaviour. The and equations represent generalized form of an overall system that what-if scenario playing and exploration of decision alternatives to approximates the peculiarity of individual elements. Conceptually, achieve organisational goals are accomplished through multiple the top-down approach considers a reductionist view [3] to simulation runs and comparing their results. understand system using the mathematical rigour from operational The rest of the paper organised as follows: the section 2 introduces research, optimization theory, and sophisticated AI algorithms. The R&I case study, section 3 illustrates the specification of the key bottom-up approach, in contrast, considers the micro-behaviour of elements of R&I organisation. The simulation runs to explore the decision alternatives of two key stakeholders of the R&I Copyright © 2017 for the individual papers by the papers' organisation namely R&I head and research project head are authors. Copying permitted for private and academic purposes. illustrated in section 4. The paper concludes in section 5 by This volume is published and copyrighted by its editors. highlighting our learnings from R&I case study and future explorations. 2 CASE STUDY DETAILS We consider an industrial Research and Innovation (R&I) organisation of an IT firm that invites new ideas from its researchers and makes appropriate attempts to convert promising ideas into innovative business offerings. The R&I organisation adopts an organisation structure and relies on a research development process to realise its goals to transform ideas into business offerings and make significant scholastic impacts. The organisation structure and behaviour are described below: (effectiveness factor of the contributed work) are associated with these research grades. Structure: The structural of R&I organisation is depicted in Fig. 1 In this setting, two key stakeholders control the organisation and its using a class diagram. As shown in the figure, the R&I organisation units. The R&I organisation is owned by a unit owner, termed as contains multiple research projects and researchers. A research R&I Head, and a research project is headed by designated project is a unit that is formed with appropriate researchers to researcher (known as Research Project Head). All research project transform a research idea into business offering. A research project heads reports to R&I head. is associated with Research Project Type wherein a research project type represents specific characteristics of the research project. For Behaviour: The process for transforming research idea into example a research project that focuses an immediate industrial business offering starts with a new idea from individual researcher problem expects quick turn-around time (for converting an idea or a group of researchers. The initiator submits new idea as a into business offering) whereas an idea that has potential to change research proposal to the research council (designated researchers) the state-of-the-art and/or state-of-the-practice takes longer for evaluation. A research project is formed once the idea is duration to reach expected maturity level. Similarly an idea that is accepted by the research council. Research project largely follows not well-explored in research community expects rigorous research process steps as described using a state machine in Fig. 2. A work, comprehensive validation strategy, and convincing research project progresses through 7 states namely research evidences to establish the success. In this paper, we consider three problem formulation/definition (RP Def), literature review (LR), research project types namely PT1, PT2 and PT3 for illustration defining line of attack (LoA), defining solution (Solutioning), purpose. PT1 type research project focuses on standard research internal technical validation through toy-yet-believable proof-of- requirements (with moderate research activity, moderate concept (TYB PoC), solution validation through near real-life solutioning activity, takes moderate time to complete and has proof-of-concept (N_R_L PoC), and external validation through medium risk), PT2 type of research project focuses on well- customer proof-of-concept (Customer PoC). An idea is transformed explored research topic (i.e., less research work, more solutioning, into business offering once the Customer PoC is completed relatively short-term and comparatively less risk), and PT3 type of successfully. The state of a research project advances based on the research project deals with long term research on unexplored topic research work contributed by researchers, and research is (i.e., more research work, more solutioning, long term and high acknowledge in research community through publications and risk). patents. For example, a research project moves from LR state to A research project comprises multiple researchers. A researcher LoA state when adequate research work is performed to address all contributes research work to the research project based on their research questions of a research project (through literature research experiences, skills and educational background. In this reviews), and the literature review outcomes are validated though case study, the researchers are classified into 4 grades called Chief appropriate publications. A research project may move from an Scientist, Senior Scientist, Scientist and Junior Researcher internal state to shelved for future work (Shelved FFW) or shelved (labelled as Developer). A range of work capability (i.e., quantum for suitable opportunity in future (Shelved PoC) state if the project of work that a researcher is capable of contributing to a research is not progressed for a specific duration. For example, the research project) and range of value weightage of the research work project in LR state may move to Shelved FFW state if the progress is not substantial for 4 weeks in a row. Essentially, each research 3 SPECIFICATION We model R&I organisation using an actor based language, named as ESL [9], that we have developed by extending the concept of actor model of computation [6] (as described in [10]) for our overarching research initiative1. ESL is capable of representing an organisation using a set of modular, autonomous and reactive actors wherein an actor may define probabilistic behaviour and interacts with other actors to support emergentism. In particular, an actor encapsulates the values that represent actor characteristics, state information, historical data, and the internal elements; actor exhibits autonomous, stochastic and temporal behaviour, and supports an interaction protocols to interact with others. The R&I organisation specification contains two types of internal actors namely research project actor and researcher actor. It also contains an actor to represent conference/workshop/journal project defines the entry-criteria and exist-criteria for all states in authority as shown in Fig. 3. terms of two factors – the progress on the core activities associated A research project actor contains: with a state (e.g., literature review activity for LR state) and a. Characteristic variables: to capture the parameters associated validation of the research work through publications. This case with state transition rules, such as the quantum of core work study uses three publication categories namely journal, conference expected for each state (as shown in Fig. 2) and expected and workshops and two sub-categories (for each category) termed publication counts for all publication categories in a state (as as tier 1 and tier 2 for defining such criteria. In addition to these shown in Fig. 2); and the other factors such as the minimum succession criteria, the R&I unit defines the rules for moving a quantum of effective work expected from an individual research project into shelved states as shown in Fig. 2. researchers to consider the work as an effective contribution The progress of a research project largely relies on the research to a research project. work contributed by individual researchers. A researcher b. State variables: to represent research project state (i.e., one of contributes work for core activities (such as literature review, the 7 states presented in Fig. 2), work progress (i.e., how much arriving at solution, and validating though PoC) and validation work is completed for core activity, publication and patent effort (publication and patent related work) based on the related work) in a state, and output produced in a state (i.e., instructions provided by the research project head. The effective number of papers accepted for different publication work contribution of an individual researcher for a research project categories). is primarily a function over effort spent on specific activity and the c. History: traces of the research project, and value weightage associated with the grade (and the quality) of a d. Internal elements: the number of researchers (having their own researcher. Further there is a non-linearity associated with the grade and behaviour) allocated to a research project. contribution from individual researcher towards the project progress. An individual contribution can be accounted to research The behaviour of a research project actor specifies the state project contribution if the contribution is above some threshold transition rules using the events that occur within research project value. For example, a researcher with a value weightage 0.5 spend actor (e.g., an event indicating an actor has completed targeted core 30 minutes in a day for literature review related work (which is work for a state or achieved specific publication targets) and/or the equivalent of 15 minutes effective work) cannot be a contribution outside of research project actor (e.g., a paper is accepted in a from an individual to a research project. One can say that the journal/conference/workshop). The behaviour of a research project minimum one hour of effective work from an individual in a day largely follows the behaviour described by the state-machine should be considered as effective work to a research project. depicted in Fig. 2 and realizes the interaction protocol depicted in The external factors, such as paper acceptance, also influence the Fig. 3. The type specific variations of the research project to research project progression. For example, state transition of a represent project types PT1, PT2 and PT3 are realised by research project is a function over number of papers accepted for a parameterising the characteristic variables. specific category. The acceptance of a paper in a A research actor encapsulates the characteristics variables that journal/conference/workshop largely depends on internal factors capture the grade, experiences, areas of interest, efficiency factors (such as the quantum of core work done, effort spent for preparing (value weightage), and the work capabilities (a list tuples a paper, the rank of the researchers who contributed to the paper describing the research activities and corresponding work limit) of and the experience of the involved researchers), and external a researcher. It also captures the work distribution instruction, i.e., factors such as the rank of the conference and inherent randomness the list of research work that a researcher should do in week (a associated with the review process, etc. researcher gets this instruction at the time of allocation to a research In this paper, we models individual elements of R&I organisation, project as Expected Work event). The state variable of a researcher i.e., different kinds of researchers, research projects and actor captures work done in week and publication counts for journal/conference/workshop authority, and their interactions to various publication categories; the history captures the experiences define R&I organisation. The specification of R&I organisation is that include the kinds of work done in the past, their quantum, and illustrated in section 3. achievements such as publication and patent histories. The behavioural specification captures inherent dynamism and uncertainty. The dynamism in work contribution from a researcher 1http://www.tcs.com/research/Pages/Model-Driven- Organization.aspx to a research project is implemented by factoring evolving value paper event to the journal/conference/workshop authority (the weightage of the researchers (value weightage changes as the events are shown in Fig. 3) when expected core work and the researcher gain experiences) and considering an uncertainty in minimum paper submission criteria are satisfied for a type of working hours for an activity (typically it is point value from publication. The authority notifies the acceptance/rejection after range). The nonlinearity in effective contribution to a research specified time delay. An accepted paper event updates research project is implemented by considering effective work (computed project state and researcher state (and history) appropriately. from the quantum of work spent in a week and value weightage of An internal state change event of a research project is triggered a researcher, where former value is uncertain and later one is once state exit criteria is satisfied. The exit criteria of PT1 research dynamic) to a research project if the effective work is significant project are shown on every transition edge in Fig 2. For example, (i.e., effective work is more than a threshold value). The difference the transition from LR state to LoA state transition is possible only in characteristics of Chief Scientist, Senior Scientist, Scientist and when 8 PW (Person Week) efforts is spent on LR activity and 2 Developer are realised by parameterising the characteristic Tier1workshop papers and 2 Tier2 workshop papers are accepted. variables of researcher actor. The external entity of this case study, i.e., journal editors, 4.1 Decision making using simulation conference organisers and workshop organisers are visualised as We illustrate relevant what-if scenarios for two stakeholders actor with probabilistic behaviour. The research project actor that of R&I organisation namely the Research Project Head and R&I sends a paper to this actor gets an acceptance or rejection Head. The goals of the research project heads are to reach business notification after a time delay. The acceptance rate and time delay offering state within desired time and make significant scholastic are pre-defined in this implementation but one may realise a contributions in terms of papers and patents. The research project complex conference system by implementing the dynamics heads explore the decision alternatives associated with researchers associated with the paper acceptance behaviour. profiles (the capability of the researchers), team distribution (research profile), the work distribution in terms of core work and 4 SIMULATION publication related work, etc. In contrast, the R&I Head, who A simulation of R&I organisation specification is essentially manages multiple research projects (with different research project execution of multiple research projects that start with RP Def state type), explores suitable strategy to maintain a steady flow of with specific number of Chief Scientists, Senior Scientists, innovative business offerings and improve research portfolio with Scientists and Developers. The simulation progresses with time high impact publications and patents. In the interest of space, we event that represents a ‘week’ time. Researchers contribute efforts discuss limited what-if scenarios in this paper. on various activities (as decided by the research project heads and the research capability of research actor) using contributed work 4.1.1 Research Project Head event (as shown in Fig. 3) every week tick. Research project In this sub-section, we first demonstrate a case (scenario 1) of an consumes contributed work event and computes effective R&I organisation with five PT1 research projects each having one contribution. Contributed effort gets wasted if effective work is below expected quantity. The research project triggers submit Chief Scientist, two Senior Scientists, four Scientists and four considers different work distribution to reduce excessive work on Developers , and then we explore improvement alternatives. patent and focus more on Tier 1 conference paper (mainly by Chief Initially, we observe the progress of R&I organisation with a setting Scientist and Senior Scientist). In this scenario, 2 research projects defined for scenario 1 by simulating R&I organisation specification have reached to customer PoC, 2 research projects have reached to for two years (considering ‘week’ as primitive simulation tick). The N_R_L PoC and 1 research project has reached to TyB PoC. For overall observation is depicted in Fig. 4.a (the graph is generated further improvement, the allocation of researchers with better by averaging 20 simulation runs to show statistically significant experience is considered (scenario 4). With this setting, 2 research result) and key data points are recorded in Table 1 (for readability projects have reached to Customer PoC, 3 research projects have purpose). As shown in the figure and table, one research project reached to N_R_L PoC. This scenario playing capability shows (out of 5) has reached to the final state (i.e., Customer PoC is how formation of simulation setting and simulation results lead to completed) whereas one research project has reached to N_R_L a decision making. In next section, we illustrate the scenario PoC state and one has reached to TYB PoC state respectively. Two playing capability of R&I head. research projects have ended up in Shelved PoC state. The key reason for slow progression is for not achieving the publication 4.1.2 R&I Head target as shown in the diagram (in Fig. 4.a). In particular, the As discussed earlier, R&I head tries to improve the flow of business research project has gone to Shelve PoC state for not contributing offerings (i.e., the ideas that reaches to Customer PoC state) and sufficient progress on Tier 1 conference paper for more than 4 maximize the publication portfolio in terms of publications and weeks. It is also observed that the effort spent on patent related patents. We configure an R&I organisation with five PT1 type activity is exceeded to an extent for some research projects. research projects, five PT2 type research projects and five PT3 type Next we demonstrate a scenario (scenario 2) that explores the research projects. The simulation results describing research impact on allocating more researchers to the earlier simulation project progresses and publication counts of these three types of settings. In this scenario, 1 Chief Scientist, 4 Senior Scientists, 8 research projects are shown in Fig. 5. As shown in the figure, the Scientists (instead of 4 in earlier setting) and 8 Developers (instead PT2 type research projects are producing more business offerings of 4 in earlier setting) are allocated to PT1 research project and the than PT3 type research projects and PT3 type research projects are progress is observed for five PT1 research projects. The producing more business offering than PT1 type research projects. progression, recorded in Table 1, is not satisfactory. The issue with In contrast, PT3 types research project are producing more tier 1 conference paper is continued in this scenario and significant publications than PT1 types research projects and PT1 type wastage on tier 1 conference related work is additionally observed. research projects are performing better than PT2 for publication. The simulation result provides a hint that the resources allocated to Thus PT2 type research project is better for churning out business the research project are not capable for tier 1 conference publication offerings but not so effective for scholastic impacts whereas PT3 i.e., researchers are putting their effort in tier 1 conference paper type research projects are better for scholastic impacts but not but the effective contribution from individual researchers are not affective for producing business offerings. adequate to reach individual threshold value. R&I head can explore the suitable combination of PT1, PT2 and We further demonstrate two scenarios – scenario 3 and scenario 4. PT3 in R&I organisation to optimise the business impact and In scenario 3, 1 Chief Scientist, 2 Senior Scientists, 4 Scientists and 4 Developers are allocated as scenario 1. However, this scenario Table 2: Simulation results representing business impacts Solutioning Definition Customer Table 1: Simulation results for research project head TyB PoC Scenario RP_Def N_R_L Comp In Progress Reached to LoA PoC PoC LR leted Shelved States Scenario 1 1 TYB PoC= 1, N_R_L PoC= 1 Shelved PoC =2 PT1=5, PT2=5, 1 0 0 0 0 0 4 11 Scenario 2 1 N_R_L PoC= 1 Shelved FFW= 1 PT3=5 Shelved PoC =2 PT1=2, PT2=8, Scenario 3 2 TYB PoC= 1 N_R_L PoC= 2 0 2 0 0 0 0 0 3 12 PT3=5 Scenario 4 2 N_R_L PoC= 3 0 PT1=4, PT2=2, 3 0 0 0 0 0 4 11 PT3=9 Table 3: Simulation results representing scholastic impacts additional expertise (other than knowing specification language) for abstracting out the system behaviour in terms of equations or Journal Tier 1 Journal Tier 2 any other aggregated form (one should specify elements as one see Conference Conference Workshop Workshop Definition them in reality), and b) exploration of decision alternative: the Scenario change specification to explore decision alternatives are localized Patents Tier 1 Tier 2 Tier 1 Tier 2 within actor (no need to find out a system parameter that represents the changes). The complete case study (not elaborated in this paper PT1=5, PT2=5, due to space limitation) makes us believe that the bottom up 1 19 30 47 139 54 112 140 PT3=5 approach, actor model and simulation are suitable for PT1=2,PT2=8, understanding the intricacy of socio technical systems. 2 15 28 47 123 52 104 118 However, we acknowledge that the case study is not sufficiently PT3=5 PT1=4, PT2=2, large to validate our claim. More experiments and real life business 3 26 55 60 130 64 163 170 critical case studies are needed. At present, we are working on a PT3=9 case study with more complexities. For example, we considered the scholastic impact through simulation runs. For example, the R&I research projects as fairly independent element and they do not head can explore what will be the situation if PT2 types of research compete with other research projects for resources (i.e., projects are encouraged within R&I organisation. We evaluate this researchers) and research outputs. Moreover, we experimented our scenario by reducing PT1 type of research project to 2 (from 5), options without any constraints such as financial constraint and increasing PT2 type of research to 8 (from 5), and keeping PT3 type resource limitation. The psychological aspect of the researchers research project count unchanged. We observed marginal while working in a research project are also not considered in this improvement in business offering as shown in Table 2 but case study. Introducing them in our case study and exploring trade- significant reduction in publications and patents counts as shown in off, competition and optimization (under constraints) are our next Table 3. For an illustration of scenario playing capability, we focus. simulated another scenario with 4 PT1 type research projects, 2 PT2 type research projects and 9 PT1 type research project. The ACKNOWLEDGEMENT simulation result is recorded in Table 2 and Table 3 respectively. We would like to thank Prof. Tony Clark (Sheffield Hallam As shown in the tables, the significant improvement is observed in University, UK) for providing necessary support to extend ESL for publication (shown in Table 3) without any trade off on business completing this case study and Prof. Balbir Barn (Middlesex offering (as shown in Table 2). University, London) for his guidance on design methodology. REFERENCES 5 DISCUSSION AND FUTURE WORK [1] Thomas, Martyn, and Frank McGarry. "Top-down vs. bottom- In this paper, we adopted bottom up approach as a design up process improvement." IEEE Software 11.4 (1994): 12-13. methodology and exploited actor model [6] as design abstraction to [2] Meadows, D.H., Wright, D., 2008. Thinking in systems: A specify R&I organisation. Further, we applied simulation technique primer. Chelsea Green Publishing to explore decision alternatives of two key stakeholders of R&I [3] Beckermann, Ansgar, Hans Flohr, and Jaegwon Kim, eds., organisation. In particular, we visualised R&I organisation using its Emergence Or Reduction? Essays on the Prospects of constituent elements; the constituent elements are specified using Nonreductive Physicalism (1992) the concept of actor; and finally emergent behaviour is observed [4] O'Connor, Timothy and Wong, Hong Yu, "Emergent through simulation run (i.e., the output produced by ESL simulation Properties", The Stanford Encyclopedia of Philosophy engine). We also demonstrated how probabilistic behaviour (e.g. (Summer 2015 Edition), Edward N. Zalta (ed.) paper acceptance), randomness (e.g. effort spent in day), non- [5] Agha, G.A., 1985. Actors: A model of concurrent computation linearity (e.g., effective research work), and dynamism (e.g. in distributed systems. Tech. rep., DTIC Document resources experience) of individual elements (that are represented [6] Hewitt, C. (2010). Actor model of computation: scalable as actors) influence the overall system behaviour (e.g. progress of robust information systems. arXiv preprintarXiv:1008.1459. research project and R&I organisation as a whole) over multiple [7] Macal, Charles M., and Michael J. North. "Tutorial on agent- simulation runs. based modelling and simulation." Journal of simulation 4.3 From methodology perspective our focus (while specifying R&I (2010): 151-162. organisation) was to find constituent elements of R&I organisation, [8] McDermott, T., Rouse, W., Goodman, S., Loper, M., 2013. understand their micro-behaviours and interactions (rather than Multi-level modeling of complex socio-technical systems. understanding the overall system behaviour of R&I organisation). Procedia Computer Science 16, 1132-1141 Moreover, the what-if scenario playing are also driven by [9] Tony Clark, Vinay Kulkarni, Souvik Barat, Balbir Barn. Actor individual elements (for example, what will be the situation if a Monitors for Adaptive Behaviour. ISEC 2017. research project head recruits more eligible researchers in a [10] Souvik Barat, Vinay Kulkarni, Tony Clark, Balbir Barn: A research project, or research head instructs team member to work Model Based Realisation of Actor Model to Conceptualise an differently) and emergentism rather the following the principles of Aid for Complex Dynamic Decision-Making. top-down approaches [11] where the primary exploration objective MODELSWARD 2017, Porto, Portugal. is to change high-level system parameters and observe system [11] Vinay Kulkarni, Souvik Barat, Tony Clark, Balbir Barn. performance. Toward overcoming accidental complexity in organisational We found the use of bottom up approach is favorable for two key decision-making. MoDELS 2015: 368-379. activities: a) specification: the specification does not expect