for his computation of agents on a regular grid structure. Here, the problem is where agents move from one grid to the next as parallel processes could result in two agents moving to the same grid cell, which is not allowed. The solution involves detect- ing where conflicts occur and using a priority ordering scheme to repeat the decision process for a lower priority agent that attempts to move onto a grid cell about to be occupied by a higher priority agent. The author makes the point that the performance of the implementation is dependent on the application2and the number of conflicts that
occur in what is a transactional based solution.
A more general introduction to the process of modelling can be found in Wilensky and Rand’s “Making Models Match: Replicating an Agent-Based Model” [WR07]. Here, the authors put forward the case for “replication” in models, citing agent-based models used in scientific papers that have never been replicated outside of the original publication. The document can be seen as a set of “best practices” for model authors, designed to ensure the replicability of the models by other researchers using different computer languages, systems and practices. The key point they make is that the rules behind the model should be explicit and open to scrutiny.
The research presented in Chapter 6 outlines an agent-based model of the London Underground and the London bus network. This was influenced by Rand’s work, “Ma- chine learning meets agent-based modeling: When not to go to a bar” [Ran06]. He suggests the idea of learning the rules of the model by observing a running system, an idea which is adapted to “learn” the operation of the London Underground by observing the data available on the real-time stream.
2.3
Analysing Tube and Bus Strikes
In chapter 8 the analysis of bus and tube strikes is used as a worked example to show the general usefulness of the real-time software developed to that point. The London Underground is used as an example by D’Lima and Medda in “A new measure of re- silience: An application to the London Underground” [DM15], where the resilience of the system to shocks is investigated. They adopt the measure of, “the speed at which a system returns to equilibrium after a disturbance away from equilibrium”, us- ing data from the real-time stream on tube positions, along with passenger flow data. 2The application used for the demonstration is Epstein and Axtell’s “Sugarscape” example in their book, Growing Artificial
The methodology works on the basis that a delay on one of the lines will lead to a sharp decrease in passenger flow. In addition, the authors take into account minor delays on connected parts of the network as a propagation of the initial shock. They take the re- siliency of the line to be a factor in a model based on Brownian motion which is fitted to the passenger flow data. In their analysis they state, “the available data was adjusted at the source to remove the effect of abnormal circumstances that may affect passenger demand such as industrial action”. They conclude with an analysis of the system under different types of shock.
On the direct results of a tube strike on the London Underground, Larcom, Rauch and Willems have published, “The Benefits of Forced Experimentation: Striking Ev- idence from the London Underground Network” [LRW15], stating that “a significant fraction of commuters on the London Underground do not travel their optimal route”. The paper then goes on to demonstrate evidence to suggest that a result of the strikes in 2014 was to force many commuters to try different routes, whereby they discovered a more efficient one. Their paper is based around the decisions commuters make in the face of imperfect information, using Oystercard travel data from Transport for Lon- don as their empirical evidence. This data is multi-modal, including tube, train, tram, Docklands Light Railway (DLR) and riverboat services, but their study only focuses on the underground data and repeat journeys made by commuters after the 48 hour strike has finished. Analysis is by fitting the data to a regression model that they propose and deriving the coefficients from the data. Toward the end of the paper, the question of travel time is more relevant to this thesis, but the authors state an obvious problem with the use of Oystercard data. Not all commuters “tap in ” and “tap out”, when they enter and exit the system3 and when changing from one tube line to another. Even when an
entry/exit record can be matched for a commuter, the path they took through the net- work can only be estimated using a shortest path, or most probable path, algorithm. The conclusions of the paper are that, “the tube network was operating so far away from its optimum, that February 2014 strike managed to improve efficiency of the system as a whole”. This is quite a bold claim to make, especially in light of the fact that the paper does not consider the effects of loading and capacity in any way. The authors finish by suggesting that strikes are quite extreme ways to encourage commuter change, but improved forms of information dissemination (e.g. “journey planner apps”) can “nudge 3Season ticket holders can simply walk through the barriers if they are open and no entry/exit data for them is collected. During
2.4. Combining Real-time and Static Data 47