An Analysis of Organizing Process of R&D Projects:Multi-agent Simulation and Case Study

The purpose of this paper is to analyze the organizing process of research projects, which affect the performance of research and development (R&D) crucially, by using agent-based simulation and case study. We develop a multi-agent simulation model that contains two types of R&D style: Japanese pharmaceutical companies’ style and Merck’s style. Simulation result proves that the senior managers observed in Merck possessing strong communication capabilities, whom we call “HWCM (Heavy-Weight Communication Manager)” in this paper, enhance initial start-up of projects. Furthermore, we study the case of Merck again and try to show the effect of HWCM on high R&D performance of Merck.


Introduction
This paper employs multi-agent simulation and case study for the purpose of analyzing the organizing process of research projects, which affect the performance of research and development (R&D) crucially. Multi-agent simulation is a relatively new method, which came to gather attention in the 1990s, yet has already been applied in various fields of studies (e.g., Epstein & Axtell, 1996;Thomas & Seibel, 1999, 2000. From existing studies up to this date, we may say that there are two types of opposite approaches in multi-agent simulation method.  Seibel (1999, 2000) have successfully improved cargo operations at Southwest Airlines by putting the multi-agent simulation method into practice.
While the two approaches are thus quite opposite, this paper will follow approach (a). We build a multi-agent model from different R&D styles at Japanese pharmaceutical companies and one of the world's largest pharmaceutical company, Merck.
We further go through the different organizational performances which the different R&D styles lead to.
In conclusion, a senior manager as HWCM Therefore, at pharmaceutical companies, release of a new drug has become of great managerial goal (Kuwashima & Takahashi, 2001).
In typical cases at Japanese pharmaceutical companies, first a research theme is set then chemists and biologists start working in development teams of one to several personnel.
Ideas are brought up within these small teams and if a potential chemical compound is discovered, chemists and biologists are added to examine various derivatives synthesized from the leading compound to attain stronger activity. Accordingly, research at individual level is launched as an official project.
As an example, we would like to investigate a R&D case of a carbapenem antibiotic Carbenin; nonproprietary name panipenem (Kuwashima, 1998 (1) At the initial stage, four teams of different approaches existed under two research themes.
(2) Since penem research team shifted to carbapenem research, synthesis methods from penem teams were brought to the joint research team.
In such major Japanese pharmaceutical companies as to rank in the top ten, generally various research themes are studied at the laboratories. In addition, multiple research teams are organized for a single theme according to different approaches.
Typically, when a certain theme or approach turns out to be promising, researchers on other themes and approaches join the hopeful one, thus an official project is launched. Relocation of researchers from theme or approach is basically agreed upon talks between researchers and locale supervisor of general manager rank.
In complete contrast, when we turn our eyes from the typical project start-up process in Japan to Heavy-Weight Communication Manager) is in presence. We intend to utilize multi-agent simulation method for our purpose.

Outline of Model
As we can see from the Carbenin case in section 2.1, "ideas" are significant at the organizing stage of pharmaceutical research projects. Therefore, we will build an multi-agent simulation model assuming that the organizing process of a research project is a "cluster formation process by agents (researchers) who possess ideas." First, we picture the R&D process in the Carbenin case at a Japanese pharmaceutical company.
For simplification, we assume that there were two themes in presence, Red and Blue, and researchers are initially involved in either. Next, as it is hard to give external criteria to the potential of themes or approaches, we simply grant that the more the ideas are in presence, the more promising the theme or approach is. Then we hypothesize that (1) agents possess ideas and move towards positions where they are more able to communicate with other ideas and agents possessing ideas; (2) when multiple clusters (research teams) exist, agents choose a cluster where they are able to communicate with more ideas.
We thus describe the rules by which agents form clusters seeking more promising research themes and approaches. We shall depict the ways that agents flow into promising clusters, thus effecting research themes and approaches to merge. (1) Path Length: L Path length is defined as following 1 to 3 and determined as in the example of Figure 1.

A path is formed between agents A and B when
A and B are the same color and all agents aligned in between A and B are also the same color. A and B are able to communicate.

2.
Path length L is determined by the number of agents between agents A and B. However, agent B is counted as well, thus when A and B is next to each other, it is L=1.
3. If multiple choices of paths exist, the shortest should be determined as path length L.
(2) Amount of Effective Idea We define the amount of an agent's effective idea as the sum of the total number of same colored agents whom the agent is able to communicate with weighted by 1/L. A typical calculation is shown in Figure 2(A). It is weighted by 1/L because we assume that where path length L is greater, the impact of a idea weakens and communication of the idea takes time.
As shown in Figure 2(A), agents belonging to a same cluster have different amounts of effective idea according to where the agent stands within the cluster. In Figure 2(A), the amount of effective idea for A is larger than B, and that for C is larger than A.
In general, agents posted nearer the center of the cluster have larger amount of effective idea. We shall describe as follows the rules of ComCom Model by the amount of effective idea.
1. An agent moves towards larger amount of effective idea.
2. An agent moves the distance of either 0 or 1 in one period.
3. An agent searches other agents within distance 1. In other words, an agent cannot perceive agents beyond distance 2.
4. An agent in contact with a cluster of other colored agents will switch sides (i.e., change color) when by doing so the amount of effective  Under above rule, an agent will prefer belonging to a cluster, and possibly a larger cluster, than staying alone.

(3) Big Agent
Big agent is a concept created with HWCM at Merck in mind, who is literally a "big" agent capable to contact and communicate directly with greater number of agents. In particular, for example, in  Figure 2(B). Accordingly, 1) a big agent has more surfaces potentially in contact with other agents, and 2) however big the agent is, path length is still measured as 1. Therefore, as in Figure 2

(4) Index
In our model, the following three indexes are indicated in numeric value and map.
1. Activity rate: This shows the number of agents that move in each period. The activity rate for period t+1 represents the amount of activity which occurred between the state in the map of time period t and t+1.

2.
Total amount of effective idea: This is the total sum of the amount of effective idea of all agents belonging to the same cluster. From the lattice model in Figure 1, we treat agents joint at a corner of a cluster as members of the cluster as well.
3. Mean cluster scale: The number of agents consisting a cluster is called cluster scale. Mean cluster scale is attained by dividing the total number of agents by the number of clusters.

Competition between Research Themes for Agents
In the present model, we presumed that every agent is involved in either Red or Blue research theme.
Here, for the purpose of investigating the effect of big agents, we would input big agents to theme Red and see if this will avail theme Red in acquiring more agents.
In practice, we compare the following two cases.
Case 1: ten L=1 Red agents; and ten L=1 Blue agents.
Case 2: eight L=1 Red agents and two L=3 Red agents, in total ten Red agents; and ten L=1 Blue agents.
Cluster formation is significantly influenced by the initial posting of agents, therefore prior to a simulation we specified a random number seed value for KK-MAS execution configuration. Particularly, in each case, random number seed value takes 1 to 30, increased by 1 in 30 trial runs. "Mean cluster scale" and "total amount of effective idea" are kept record up to period 300 in each trial run.
First we compare the results at period 300 in the   Thus, big agents choose between themes as well as regular agents. In other words, the theme big agents are assigned to at an initial stage is not critical when research themes compete to acquire more agents, in other words, when they compete for research resources.

Mean Cluster Scale and Total Amount of Effective Idea
Based on above finding, next we shall give the statistics on cluster scale and total amount of effective idea on the whole disregarding themes Red and Blue. Figure 3 shows the average mean cluster scale for 30 trial runs concerning Case 1 and Case 2.
Accordingly, Case 2 where big agents are cast in, reveals a tendency to experience a growth of mean cluster scale at an earlier stage. Nevertheless, at period 300, mean cluster scales are slightly more than 6 in either case, thus the gap is narrowed.
Therefore, big agents do not commit to the size of clusters but to the earlier formation of clusters.
Likewise, in Figure 4, total amount of effective idea for each period in 30 trial runs are averaged out.
As the figure indicates, total amount of effective idea increases at an earlier stage in Case 2 where big agents are cast in, similar to the analysis of mean cluster scale. However, at period 300, total amount of effective idea in either case are slightly more than 100; again the gap between Case 1 and Case 2 are narrowed.
In sum, 1) input of big agents as a whole enhance the formation of clusters and total amount of effective idea at an earlier stage. 2) Though speed is enhanced, neither mean cluster scale nor total amount of effective idea take lager values in the end. The reason that our study conducted case analysis in addition to simulation analysis was to provide complementary investigation to the result drawn from the "rather abstract simulation model (Section 1, Approach (a))"; that is, to see how the influences of HWCM on organizing processes are observed in actual pharmaceutical product development. While typically such products as automobiles are produced under highly successful routine R&D and product development, pharmaceutical product R&D projects, which is the object of our analysis, operates under such sparing success rate of thousands to ten thousands to one:

HWCM in Practice and R&D
Projects literally emerge from a single idea. When we analyze such cases as this one, empirical measurement and analysis of a certain variable (in this paper the presence of HWCM), given that other conditions are fixed, is unreliable. We believe that combining simulation and case analysis, as we have done in this paper, proves as a useful research approach when analyzing an innovation of low success rate.
Besides, it is possible to regard this kind of research method, combining simulation and case analysis, as a stepping stone from simulation to further empirical study. Our procedure was to build a simulation model inspired by the case at Merck, then based on the simulation result, go back to the case to analyze the HWCM effect. From the results of this paper, our future object is to put more emphasis on case analysis; We aim to measure HWCM by manipulated profiles and behavioral patterns along with more systematic analysis of the relation between HWCM and R&D performances.