Modeling R&D Strategy as a Network Search Problem
The principal reason for the relatively poor performance of depth-first search is simply that the algorithm finds root nodes (nodes without parents) and these constitute a cul de sac from which the agent has no means of escape. A richer model would allow agents to recognize root nodes and give them some strategy for starting their searches from some other node in the network. A larger scale of network might also yield qualitatively different results.
It is generally well recognized by artificial intelligence scientists that a depth-first search is relatively more efficient (and a breadth-first search relatively less efficient) when the search network is deeper and has fewer points of entry. This is what we observe in our simulations since, in the 10-leaf-node network more nodes are acquired by agents adopting depth-first search and fewer nodes are acquired by agents adopting breadth-first search than is found for the 100-leaf-node network.
The greater efficiency of search over technology purchase also follows from the setup specification.
Every node that is acquired reduces the unit production cost of the agent. Agents that only search for technologies must find acquire a larger number of nodes than agents that buy their node-acquisitions. We did not, in these simulations, allow for agents that bought technology to use the resources not allocated to R&D to engage in other activities which might have enhanced their financial positions. We do not, therefore, suggest that our simulation experiments suggest anything very compelling about the virtues of technology licensing vis a vis in-house technological development.
Generated with CERN WebMaker