|
The Future of Parking |
|
The amount of parking related data available nowadays is rising at an incredible pace. This parking information is generated by an always increasing number of purposely allocated sensors and cameras, as well as collected from smartphones, GPSes and smart cars. Unfortunately, this data presents two major drawbacks: it is noisy and partial.
These characteristics have a severe impact on the current applications of this data, which lies for the most part unused. Boetho exploits this untapped potential, via a purposely designed machine learning algorithm able to cope with the mentioned difficulties. Indeed, despite the limitations of the data sets involved, our algorithm is able to produce robust inferences on future parking occupancy, with only past occupancy as input. Once the future occupancy is uncovered, several optimization problems can be solved. For instance, we already know how to achieve optimal dynamic pricing, booking system and allocation. Looking at a more distant future, we envision Boetho’s products to play a key role in the emerging market of self-driving cars, smart cities and IoT. Indeed, the technical framework we have developed is ready for this future environment. |
A Tailor-made Bayesian Algorithm
Why is there currently no effective use of parking data?
Obtaining a continuous flow of accurate parking data is expensive and not economically viable. Information captured by cameras and sensors is affected by the limited quality of the hardware, while information acquired by Parking Apps, GPSes or cars is only partial, as no company has access to the entire customer base. When analysing data sets, the common practice is to utilise frequentist probabilistic models, which fail to perform well on these type of data sets. Indeed, proper calibration of such models requires several years of records. Demanding this much data is expensive, time consuming and creates models too dependent on the past. Boetho’s solution is to adapt a Bayesian point of view. Why use a Bayesian approach? Bayesian models continuously incorporate observed data into the model. This dependence on new data provides the perfect base to deal with noisy and partial data sets. Indeed, in the context of parking data, Boetho’s predictive algorithm cleverly uses state-space DLM tools to generate robust results with only a few months of data. On top of that, Bayesian models are inherently adaptive, meaning that they can easily pick up changes in the data due to any new emerging trend. Graphics description The graphics showcase the predicted occupancy generated by our algorithm using a combination of a weak-ahead prediction combined with an online learning component (in black) against the real-life data (in red) for both a day and a week ahead scenario. |
Optimal Allocation
The need for a global network optimisation
Predicting the parking occupancy within a given street or optimizing the price for a parking location is one thing, trying to optimally allocate and price the parking resources of a whole city or of a network of parking locations according to personalized utility functions for all drivers is an entirely different problem. To solve this problem on a city scale the algorithm needs to be ready to adapt to the drivers’ response to the new information introduced by the system. For instance, if an app shows a low occupancy area next to a busy one, all the rational drivers having access to this information will go to the low one thus translating the problem without solving it. Indeed considering an independent problem for each driver does not solve the global multi-dimensional problem taking into account the matrix of cross-relationships between each of them. Furthermore, the size of a city brings additional computational limits requiring specific tools to solve this sparse-matrix optimisation in linear time. Boetho’s solution A utility function is associated to each driver taking as input a vector with different preferences. The combination of those utility functions, Boetho’s inferences and dynamic pricing creates a large scale quadratic Lagrangian. Boetho’s allocation algorithm can solve this optimisation problem efficiently, obtaining the global Nash equilibrium. This solution optimises the satisfaction of every driver and the use of the parking resources of the whole city, while at the same time provides the optimal pricing strategy for the network of parking locations, thus maximizing the profitability of this network. While the city scale application of this approach is distant, this allocation optimisation can be used on a network, or combination of independent networds of parking locations to optimise the profit of the whole group of networks and not only one specific asset. Graphics description The graphics uses occupancy rates of the street of San Francisco to simulate a city. Each bar represents a parking location, the higher and yellower the higher the occupancy of that location is. The top graphic shows the evolution of the occupancy rate for a subset of streets from San Francisco during 24h. The graphic below it represents the same data with Boetho’s optimal allocation algorithm implemented. The implementation includes dynamic pricing thus a lower but more uniform occupancy means better cash flow for the parking. In addition the graphics clearly show the ability of the optimal allocation algorithm to solve the driver and societal challenges of parking related problems on a city scale. Indeed, the second graph presents on average lower bars, which imply lower traffic jams, pollution and time wasted in the search of a parking spot. |