Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Dynamic Influence on Replicator Evolution for the Propagation of Competing Technologies

Dynamic Influence on Replicator Evolution for the Propagation of Competing Technologies This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 Dynamic Influence on Replicator Evolution for the Propagation of Competing Technologies Elijah D. Bolluyt, Comaniciu C., Member IEEE factors on individualization of strategies [3][10]; however, Abstract— This work introduces a novel modified Replicator while including some environmental dynamism, these Dynamics model, which includes external influences on the approaches still do not model an environment in which population. This framework models a realistic market into which external parties actively influence the market over time, while companies, the external dynamic influences, invest resources in the individuals in the population remain homogeneous. order to bolster their product’s standing and increase their Additionally, previous applications of the replicator market share. The dynamic influences change in each time step framework to realistic scenarios have not focused on dynamic of the game, and directly modify the payoff matrix of the population’s interactions. The model can learn from real data modification over time to the payoff matrix [9][11][12][13]. In how each influence affects the market, and can be used to addition, to the best of our knowledge, no work exists in simulate and predict the outcome of a real system. We integrating real data into such a model in order to make specifically analyze how a new technology can compete and predictions on a real world system. These characteristics are attempt to unseat an entrenched technology as the market leader. necessary to create a model that can accurately emulate and We establish a relationship between the external influences and predict a realistic marketplace. the population payoff matrix and show how the system can be implemented to predict outcomes in a real market by simulating This work introduces a general method to realistically the rise of the Android mobile operating system over its primary model the dynamic influence of external factors on an competition, the iPhone, from 2009-2017. unmodified population, and investigates the actions necessary for such an outside influence to achieve specific goals. We Index Terms— Evolutionary game theory, Product markets, then apply this method to specifically address the case of a Replicator dynamics market that is initially dominated by an entrenched technology, describing the conditions necessary for a smaller, I. INTRODUCTION emerging technology to compete and, eventually, overtake the VERY real environment is in a state of perpetual flux, market leader. Our model not only predicts market conditions especially those that share a close relationship with based on input information, but it also provides the user with technology. As technologies evolve in a free market, information about the intrinsic properties of the marketplace. competition inevitably arises; sometimes, one company Because our approach emphasizes realism and applications dominates, and other companies expend resources to make for modeling real markets, we conduct a simulation of the real inroads. No matter the specifics, competition leads to each case of the world’s smartphone market from 2009-2017 to individual in the market choosing which technology to use prove the efficacy of our model. This range captures the rise based on the relative value of each to the user. of the Android smartphone operating system (OS) and its Replicator dynamics provides a useful framework in which ascension to dominance over the Apple iPhone, which had to analyze competitive markets, as it models a population of hitherto led the market [4]. identical individuals who choose a strategy based on its projected payoff, or the value to the user [1]. However, the II. MODEL original replicator equation’s payoff is a function only of the A. General Replicator Dynamics Model portion of the population which employs each strategy; the strategy with the larger user base dominates [1]. In a free and Replicator dynamics models the evolution of a population of active market, however, each rational company will invest identical individuals who employ one of n strategies and assets to increase its product’s adoption rate, and thus its receive a payoff from each interaction with other individuals based only on each individual’s choice of strategy [1]. A profits. player employing strategy i who interacts with one employing Previous extensions of the replicator dynamic model have a strategy j receives a payoff 𝐴𝐴 ; the set of payoffs can be included feedback loops between the environment and the expressed as a matrix: strategy payoffs [2], as well as the effects of localization 𝐴𝐴 𝐴𝐴 ⋯ 𝐴𝐴 11 12 1𝑛𝑛 Manuscript submitted February 12, 2018; revised June 24, 2018 and 𝐴𝐴 𝐴𝐴 ⋯ 𝐴𝐴 21 22 2𝑛𝑛 September 10, 2018. 𝐴𝐴 = � � (1) ⋮ ⋮ ⋱ ⋮ E. Bolluyt and C. Comaniciu are with the Department of Electrical and Computer Engineering, Stevens Institute of Technology, Hoboken, NJ 07030 𝐴𝐴 𝐴𝐴 ⋯ 𝐴𝐴 𝑛𝑛1 𝑛𝑛2 𝑛𝑛𝑛𝑛 USA (e-mail: ebolluyt@stevens.edu; ccomanic@stevens.edu). Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. 𝑖𝑖𝑖𝑖 This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 A fraction 𝑥𝑥 ∈ [0,1] of the population employs the strategy 𝑥𝑥 ̇ > 0 → 𝐴𝐴 𝑥𝑥 +𝐴𝐴 𝑥𝑥 > 𝐴𝐴 𝑥𝑥 +𝐴𝐴 𝑥𝑥 (6) 𝑖𝑖 1 11 1 12 2 21 1 22 2 i. The replicator equation expresses the growth rate of this portion of the population: When combined with (5), we see that growth of 𝑥𝑥 requires: 𝑥𝑥 >�𝑥𝑥 (7) 1 1 𝑇𝑇 𝑥𝑥 ̇ = 𝑥𝑥 ((𝐴𝐴 𝑥𝑥 ) −𝑥𝑥 𝐴𝐴 𝑥𝑥 ) (2) 𝚤𝚤 𝑖𝑖 𝑖𝑖 This condition implies that the system moves away from this mixed equilibrium, not towards it, making this point an A strategy’s usage only increases if its payoff is higher than unstable equilibrium point; it also tells us that the payoff the average payoff in the population. Thus, the expected matrix A must favor a low value mixed equilibrium in order payoff of each strategy directly modifies the individuals’ for 𝑥𝑥 to continue to grow. choices of which strategy to employ. Condition (7), combined with (2) and (6), allow us to The growth expressed by the replicator equation (2) analyze the effects and desirable traits of each of the four captures the change in the fractions of the population that payoff elements from the standpoint of strategy 1’s proponent, employ each strategy over time, showing how a higher payoff the external influence working in favor of strategy 1. causes a strategy to gain prominence in the population. should be increased according to (2). To attain growth, 𝐴𝐴 Note that for the population’s strategy distribution to change This compensates for the small initial value of 𝑥𝑥 . As 𝑥𝑥 1 1 significantly, the elements of A must be very different. It can approaches dominance, 𝐴𝐴 can be allowed to decrease as be seen from (2) that if all elements of A are approximately lower growth is expected. 𝐴𝐴 should be kept large in the equal, then the payoff of each strategy will be very close to the beginning to allow for the dominance of strategy 1, but its average, and thus no strategy’s growth rate will be significant. value becomes less important as 𝑥𝑥 increases. 𝐴𝐴 should also 1 21 B. Two Strategy Replicator Dynamics Model be kept large enough to allow for dominance of 𝑥𝑥 , but it should decrease as 𝑥𝑥 increases as this will allow the growth To model a technology market with one entrenched, rate of 𝑥𝑥 to increase. dominant technology and one new technology, we formulate a Note that 𝐴𝐴 is not under the control of an influence of population with two strategies available to each individual, 22 strategy 1. A technology’s interactions with itself should be where each strategy represents the usage of one of the two controlled only by the company that owns it. We can expect, technologies. The payoff matrix A becomes: however, that in a population where 𝑥𝑥 is increasing, 𝐴𝐴 is 1 22 likely to be small in magnitude. 𝐴𝐴 𝐴𝐴 11 12 𝐴𝐴 = � � (3) 𝐴𝐴 𝐴𝐴 21 22 C. Dynamic Influence Extension We now formulate an extension to the payoff matrix that 1) Equilibrium Conditions allows such modifications, as outlined above, to be made by The two strategy replicator dynamics model has three an external influence on the market. We construct a column equilibrium points at which the growth of each strategy is vector where each element 𝑦𝑦 ,𝑖𝑖 ∈ [1,𝑛𝑛 ] represents an zero: 𝑖𝑖 𝑦𝑦 𝑥𝑥 ̇ = 𝑥𝑥 ̇ = 0 (4) influencing factor. These factors are inputs that affect the 1 2 market and can be modified by the policies of companies in Two are trivial, wherein one strategy grows to engulf the the market, such as product pricing or investments (which entire market and the other is driven to extinction. The third directly impact the quality of the product). The number of the equilibrium point can be found by using (2) and (4): inputs 𝑛𝑛 needed to accurately characterize the markets will 𝑦𝑦 depend entirely on the properties of the markets under 𝐴𝐴 −𝐴𝐴 22 12 �𝑥𝑥 = (5) 1 investigation, and has no direct relation to 𝑛𝑛 . The system will (𝐴𝐴 +𝐴𝐴 )−(𝐴𝐴 +𝐴𝐴 ) 11 22 21 12 be able to model the market evolution only if the major relevant inputs that characterize that particular market are It is important to note that the equilibrium conditions reached are dependent not only on the system’s parameters, but also on identified. the initial conditions; if one strategy starts with a market share Considering n influencing input factors, each element of of zero, then it never grows and the system stays constant. the payoff matrix becomes a linear combination of these factors. The linear coefficients are expressed as a newly 2) Impact of Payoff Matrix Element Values on System defined 𝑛𝑛 ×𝑛𝑛 matrix 𝛼𝛼 . To illustrate the influence of the 𝑦𝑦 Evolution input parameters on the elements of matrix A, the payoff The elements of the payoff matrix directly influence the matrix A can be re-indexed with a single index 𝑘𝑘 ∈ [1,𝑛𝑛 ], outcome of the market. Thus, the behavior of the system can ( ) traversing row by row, where 𝑘𝑘 = 𝑛𝑛 𝑖𝑖 − 1 +𝑗𝑗 and i,j are the be changed by modifying the elements of the payoff matrix in original indices 𝑖𝑖 ,𝑗𝑗 ∈ [1,𝑛𝑛 ]. Consequently, each element of specific ways. An external party looking to influence the the payoff matrix can be expressed as a linear combination of market should attempt to make its strategy (i.e. technology) the input factors: constantly grow and finally reach equilibrium at a dominant 𝐴𝐴 = 𝛼𝛼 𝑦𝑦 (8) position. 𝑘𝑘 𝑘𝑘 where 𝛼𝛼 is the kth row of the coefficient matrix 𝛼𝛼 . In order to grow, a strategy’s payoff must be greater than 𝑘𝑘 This representation allows the external influences to have a the average; (2) indicates that this condition is necessary for direct impact on the outcome of the market. As stated above, the growth rate to be positive. This constraint leads to a few constraints must be placed on this matrix 𝛼𝛼 to maintain condition (6) Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 realism; specifically, any factor controlled by company i must only affect terms involving strategy i. This modification makes A entirely dynamic, able to take on different values during each iteration of the simulation. This allows the model to take into account the influence of external parties which change their strategies based on the market. An external party can evaluate how its input affects the payoff matrix and thus the market, and react by changing its input values accordingly. These factors may include investments a company makes into its product, pricing of its product, or any other values that have a direct impact on the quality and availability of its technology. The value of 𝛼𝛼 is a constant, and it represents an intrinsic property of the market’s interactions with the input factors. For any particular market and input set y, 𝛼𝛼 must be learned using real data. The value of 𝛼𝛼 gives the user insight into the market, showing the influence of each input on the utility of Fig 2. System outputs. The inflection point in the market and the values in each product. This stands in contrast to other learning the validation points are accurately predicted by the system. methods, in which the learned parameters give no insight into the simulated system. necessarily reach this equilibrium, however, because the In our simulation, we employ a simple exhaustive search system changes over time. The system converges to a final and evaluate candidate results via a least mean squares error equilibrium only when the inputs y, and thus the payoff metric, while withholding 20% of the available data in matrix, converge to a constant value. accordance with common practice [5]. When training the system, learning the parameters of A, this last 20% of data was III. SIMULATION withheld from the set; this tests the model’s predictive A. Setup and Data capability. A mean squared error metric was used as it is To validate the accuracy of the proposed model for practical computationally simple and accurately represents the validity applications, a simulation was conducted on data from the of the result. global smartphone market from Quarter 1 (Q1) 2009 to Q1 𝑡𝑡 2 𝑓𝑓 𝐽𝐽 = �𝑥𝑥 (𝑡𝑡 )−𝑥𝑥 (𝑡𝑡 )� (9) 1 1 𝑡𝑡=1 2017. This range was chosen as it captures the rise of the 𝑡𝑡 𝑓𝑓 Google Android platform to dominance over the Apple iPhone Where 𝑥𝑥 (𝑡𝑡 ) is the prediction at time 𝑡𝑡 and 𝑥𝑥 (𝑡𝑡 ) is the real in market share. In order to simulate a two-strategy system, data at time 𝑡𝑡 . the smartphone market is simplified to just Android and iOS; Due to the time dynamic nature of these modifications, the while other platforms existed, these two grew to dominate the system no longer has a single guaranteed equilibrium value to market as the modern wave of touchscreen smartphones which it will converge. At each time step, the system has a developed [4]. target equilibrium to which it would converge if the system The system takes four input factors 𝑦𝑦 : average pricing for 𝑖𝑖 inputs y were to stay constant. The system does not Fig 1. System inputs. These inputs contribute directly to the growth or Fig 3. Payoff Matrix Coefficients. These four computed factors determine shrinkage of each technology in the market. Note the difference in the market behavior as discussed in Section II, and their values are a direct magnitudes between the investment factors, as well as the opposing trend result of the system inputs. Note the small magnitude of A as well as the directions of the products’ pricing. trend directions of A , A , and A . 11 12 21 Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 TABLE I OPTIMIZED VALUES OF 𝛼𝛼 𝒚𝒚 𝒚𝒚 𝒚𝒚 𝒚𝒚 𝟏𝟏 𝟐𝟐 𝟑𝟑 𝟒𝟒 𝑨𝑨 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝛼𝛼 11 12 13 14 𝑨𝑨 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝟐𝟐𝟏𝟏 21 22 23 24 Simulation 𝑨𝑨 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝟏𝟏𝟐𝟐 31 32 33 34 𝑨𝑨 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝛼𝛼 41 42 43 44 Real Data 4 0 -1 0 0 3 1 3 3 0 3 1 0 4 0 -1 Constant Market 4 0 2 0 (iOS dominant) 0 3 -3 4 3 0 4 -3 0 4 0 2 The value structure sho ws to which input 𝑦𝑦 and payoff matrix element 𝑖𝑖 𝐴𝐴 each element of 𝛼𝛼 belongs, as each is a linear coefficient for one input in one payoff element. 𝑦𝑦 : Android Investment 𝑦𝑦 : Android Price 1 3 𝑦𝑦 : iOS Investment 𝑦𝑦 : iOS Price Fig 4. System outputs in simulation with constant inputs, using 𝛼𝛼 optimized 2 4 with real data. This shows that the strategy Android employed to gain dominance in the market had a great effect on increasing its final market C. Conditions on 𝛼𝛼 for Optimization share. A is normalized to keep all elements 𝐴𝐴 ∈ [0,1], in Android and iOS smartphones sold each quarter [8], and accordance with the standard replicator equation conditions quarterly research investments in mobile from Apple [6] and (1). Thus the magnitude of 𝛼𝛼 does not matter, as the system’s Samsung [7], the largest Android phone manufacturer. These behavior is determined by the relative size of its elements. factors were chosen as they are determined by the device The grid search was conducted over an interval [−𝑟𝑟 ,𝑟𝑟 ] by manufacturers’ direct influence and have a direct impact on integer steps; the search domain was optimized as a hyper the devices’ market performance. The values of these factors parameter of the system over multiple trials, and a value of over the simulation period are shown in Fig. 1; these values 𝑟𝑟 = 4 was chosen for this simulation. The complexity of the were obtained directly from real data [6][7][8]. The data was 𝑛𝑛 ∗𝑛𝑛 /2 𝑦𝑦 search increases on the order of (2𝑟𝑟 + 1) , or (2𝑟𝑟 + 1) sampled quarterly, to match the available frequency of in this particular simulation; this makes increasing the search investment data from the companies’ quarterly reports. This interval very computationally expensive. The value of 𝑟𝑟 gave a total of 33 time samples in the simulation period. Even determines the resolution of the resulting 𝛼𝛼 , and must be with this relatively sparse data, the model remained effective; chosen by trial and error. For our simulation we started with results are discussed further in Section III-D. 𝑟𝑟 = 1 and increased 𝑟𝑟 until the error between the simulated The system is initialized at the market shares determined by −5 curve and the real market evolution was less than 4 × 10 . inputting the first real data point into the system; this is vital to Since the search is conducted by integer steps and the obtaining a result that mirrors the reality of the market. The resulting A is normalized, an interval of [−𝑟𝑟 ,𝑟𝑟 ] is equivalent initial conditions are central to the determination of the to searching on an interval of [−1,1] with a step size of 1/𝑟𝑟. A system’s behavior, as mentioned in Section II-B, so they must higher value for 𝑟𝑟 is desirable for better resolution of the be based on the real data. search, however a high value for r induces a higher As discussed in IIC, a subset consisting of approximately computational complexity. The complexity increases with 20% of the latest data, or 7 time samples, is withheld from the both the number of shares in the market, 𝑛𝑛 , and the number of training set to serve as a validation set in order to determine input factors 𝑛𝑛 . This model is well suited to technological the system’s predictive capability. 𝑦𝑦 markets, where typically only a few companies have the B. Implementation capability to compete effectively. Technological markets also often have a high barrier to entry, making the study of the A grid search is employed to optimize 𝛼𝛼 . For each candidate confict between a dominant product and a new competitor value, the complete evolution of the system over the training especially relevant. interval is computed. For each time step in the evolution, the Imposing practical, common sense constraints, the effect of matrix A is computed as defined in (8) and normalized, and the growth rates of the two technologies (Android and iOS) the companies’ influences on their respective products must be identical; a change in an input parameter for product 1 must are computed (2). The market values at all time-steps are computed using the A and growth rate obtained for the cause a variation in the payoff coefficients symmetric to the result of the same change in the same input parameter for candidate 𝛼𝛼 value, and the error J between this prediction and product 2. If parameters 𝑦𝑦 and 𝑦𝑦 are the same input for the the real data is computed. The value of 𝛼𝛼 with the lowest error 1 2 two respective products, then the rates of change of A’s over the training interval is chosen. This 𝛼𝛼 is used for elements must be constrained by validation in section D. Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. 𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 𝟐𝟐𝟐𝟐 𝟏𝟏𝟏𝟏 This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 𝜕𝜕𝐴𝐴 𝜕𝜕𝐴𝐴 E. Analysis and Significance of 𝛼𝛼 { } = ;𝑖𝑖 ,𝑗𝑗 = 1,2 (10) 𝜕𝜕𝑦𝑦 𝜕𝜕𝑦𝑦 𝑖𝑖 𝑖𝑖 The value of 𝛼𝛼 obtained from the simulation using real data Since these partial derivatives are linear, this can be is shown in Table 1. The values of its individual elements accomplished by constraining the values of the elements of 𝛼𝛼 show the effect of each input on the payoff matrix, and thus which correspond to these enforced relations to be equal the market. As can be expected, higher investment of each because the elements of 𝛼𝛼 are the rates of change of A’s company in its own product positively effects its own elements with respect to the input parameters y. These and α ), and higher pricing of each product’s utility (α 11 42 constraints ensure that the system remains symmetric. If we product negatively effects its utility (α and α ). Similarly, 13 44 linearly index A with a single variable 𝑙𝑙 , traversing row by row increased research spending also increases the utility that the 𝜕𝜕𝐴𝐴 𝑙𝑙 as stated in Section IIC, corresponds to 𝛼𝛼 as each row of product’s users obtain from interacting with the competitor 𝑙𝑙 𝑘𝑘 𝜕𝜕𝑦𝑦 𝑘𝑘 technology (α and α ), indicating efforts to increase cross 𝛼𝛼 corresponds to one element of A, as expressed in (8). 22 31 compatibility to increase market uptake. Note that any influence on technology i should not exert an A second simulation was run, in which the optimized value influence on 𝐴𝐴 , as discussed in Section IIB. Thus the 𝑖𝑖𝑖𝑖 of α was used to predict a market outcome where the elements of 𝛼𝛼 corresponding to these factors are zero. In a companies kept a constant level of investment and pricing, two strategy population, there will be one such factor for each equal to that in the first time step. This was done to assess the input as each input corresponds exclusively to one technology. effectiveness of the strategies employed by each company, and D. Analysis determine what effect the feedback from the market had on the outcome. The results of this simulation can be seen in Fig. 4; The simulated and real data can be seen in Fig. 2. The Android still becomes dominant, but its market share saturates model was trained using the training set, and then the system at a lower value. This would seem to indicate that Android’s was allowed to evolve on its own, employing its learned value strategy of decreasing price and increasing investment payed of α and the real market input data (prices and investments) off significantly. for the duration corresponding to the validation set. Its Finally, a third simulation was conducted in which the computed market share values had a mean squared error −5 market shares were held constant. This was done to observe of 3.0 × 10 with respect to the real data for this period. what type of market would have allowed iOS to remain This reflects the ability of the system to predict future data dominant, given the real input data. The value of α optimized with high accuracy. While this simulation approached an for this hypothetical situation is also shown in Table 1. equilibrium state, this is due to the inputs remaining fairly Investments play a similar role in increasing a product’s constant near the end of the simulation period; such a utility, but higher prices now have a large positive effect on convergence is not guaranteed in this model, as discussed in the utility of each product (α , α ). Such an unrealistic Section II-C. 13 44 property makes sense because iOS devices have a much higher The values of A’s elements (Fig. 3) in the simulation price compared to Android devices throughout the simulation conformed well to expected behaviors from Section IIB. 𝐴𝐴 period. If consumers wanted to pay a higher price, then this increases while technology 1 (Android) holds a small market would have improved iOS’s market share size rather than share, and starts to decrease after the inflection point in the shrinking it. market. 𝐴𝐴 remains large with some fluctuation throughout Analyzing these simulations shows that the properties of α the simulation. 𝐴𝐴 starts large to allow for initial growth, reflect real properties of the market. This direct relation is the then decreases as Android reaches a dominant market position. core advantage of this model over other learning based 𝐴𝐴 is small in magnitude in comparison to the other methods: the learned parameters give direct insight into the coefficients; this is to be expected in a market where market reaction relative to the inputs. This could potentially technology 1 is growing its market share. However, 𝐴𝐴 did allow a market player to not only simulate strategies and increase over the simulation period, reflecting Apple’s observe the outcome, but also to understand how the increased investment in iOS and their pricing decreases near individual inputs under its control could modify the market the end of the observed period, likely an effort to grow iOS’s share evolution. The players could then devise a strategy that market share. achieves the desired outcome based on this knowledge of the The system’s accuracy to both the real market values and market. the expected internal behavior indicate its efficacy in modelling the marketplace. The small input data size of 33 IV. CONCLUSION time steps coupled with these results indicate that the system is able to perform with much less data than would be required This work extends the Replicator Dynamics framework to for many learning based models. Its replicator foundation analyze a system with external influencing factors. It modifies specifically models strategy fitness as a function of the payoff matrix of the core system to adapt to external population, an important feature when dealing with factors, and learns parameters to model that dependence. This technologies that require a user base to stay viable. Versus a enables the system to contend with external factors that general learning method, the model only has to learn the influence an otherwise symmetric population, an ability not specific nature of a particular market rather than learning the present in previous replicator formulations. This modified nature of markets in general; this allows it to perform well framework was applied to the rise of the Android platform to with less data, as less information has to be learned. mobile OS market dominance between 2009 and 2017; the proposed system accurately modeled the evolution of each Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. 𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 technology’s market share for several fiscal quarters, [11] J. Huang, J. Wang, H. Zhang, and N. Wang, “2017 employing parameters obtained via optimization over a Ieee 2nd Advanced Information Technology, training set. The simulation also produced intermediate Electronic and Automation Control Conference parameter behavior which closely corresponded with (Iaeac).” Network Defense Strategy Selection Based theoretical predictions for a scenario such as the rise of on Best-Response Dynamic Evolutionary Game Android over iOS, starting from a small initial market share. Model, IEEE, 2017, pp. 2611–2615. This simulation result indicates that the system can be [12] S. D’Oro, L. Galluccio, S. Palazzo, and G. Schembra, employed to accurately model and predict the evolution of “A Game Theoretic Approach for Distributed markets in real-world scenarios. This model could aid in Resource Allocation and Orchestration of predicting the outcome of variations in strategy of companies Softwarized Networks.” IEEE Journal on Selected in a marketplace, making it a useful tool to create new Areas in Communications, vol. 35, no. 3, 2017, pp. strategies which can produce a desired outcome. 721–735. [13] M. A. Shattal, A. Wisnieswska, A. Al-Fugaha, B. Khan, K. Dombrowski, “Evolutionary Game Theory REFERENCES Perspective on Dynamic Spectrum Access Etiquette.” [1] R. Cressman, “Symmetric Normal Form Games,” in IEEE Access, (2017 08 22), 2017. Evolutionary Dynamics and Extensive Form Games, Cambridge, MA, USA: The MIT Press, 2003, ch. 2, Elijah D. Bolluyt received the B.S. degree sec. 1, pp.19-22. in engineering physics and the M.E. degree [2] J.S. Weitz, C. Eksin, K. Paarporn, S. P. Brown, and in electrical engineering from Stevens W. C. Ratcliff, “An oscillating tragedy of the Institute of Technology, Hoboken, NJ, in commons in replicator dynamics with game- 2017. He is currently pursuing the Ph.D. environment feedback,” PNAS, vol. 113, no. 47, pp. degree in computer engineering at Stevens E7518-E7525, Nov. 2016. [Online]. DOI: Institute of Technology, Hoboken, NJ. 10.1073/pnas.1604096113. During the summer of 2015, he was a [3] F. Pieper and S. Mostaghim, "Influence of dynamic Scholars summer researcher with the Department of Physics environments on agent strategies," 2016 IEEE and Engineering Physics at Stevens Institute of Technology. Symposium Series on Computational Intelligence During the summer of 2016, he was an Engineering Intern at (SSCI), Athens, 2016, pp. 1-8. doi: 10.1109/SSCI. Kulite Semiconductor, Inc. in Leonia, NJ. His current 2016.7850159 research interests focus on the development of computational [4] “Mobile Operating System Market Share frameworks, employing game theory and machine learning Worldwide,” StatCounter Global Stats. Accessed techniques. September 15, 2017. [Online]. Available: Mr. Bolluyt is a recipient of the Provost’s Doctoral http://gs.statcounter.com/os-market-share/mobile/ Fellowship (Stevens Institute of Technology). worldwide/ [5] T. M. Mitchell, “Decision Tree Learning,” in Cristina Comaniciu received the M.S. Machine Learning. New York, NY, USA: McGraw- degree in electronics from the Polytechnic Hill, 1997, ch. 3, sec. 7.1, pp.68-70. University of Bucharest in 1993, and the [6] “Earnings Releases,” Apple Inc., Investor Relations. Ph.D. degree in electrical and computer Accessed September 15, 2017. [Online]. Available: engineering from Rutgers University in http://investor.apple.com/financials.cfm [7] “Earnings Releases,” Samsung Global, Investor From 2002 to 2003 she was a Relations. Accessed September 15, 2017. [Online]. postdoctoral fellow with the Department Available: http://www.samsung.com/global/ir/ of Electrical Engineering, Princeton University. Since August financial-information/earnings-release/ 2003, she is with Stevens Institute of Technology, Department [8] F. Richter, “The Smartphone Price Gap,” Statista, of Electrical and Computer Engineering, where she is now an June 2, 2016. [Online]. Available: Associate Professor and serves as Associate Department Chair https://www.statista.com/chart/4954/smartphone- for Graduate Studies. In Fall 2011 she was a visiting faculty average-selling-prices/ fellow with the Department of Electrical Engineering, [9] K. Coninx and T. Holvoet, “Darwin in Smart Power Princeton University. She served as an associate editor for the Grids – Evolutionary Game Theory for Analyzing IEEE COMMUNICATION LETTERS (2007-2011). Self-Organization in Demand-Side Aggregation,” Professor Comaniciu is a recipient of the 2007 IEEE International Conference on Self-Adaptive and Self- Marconi Best Paper Prize Award in Wireless Communications Organizing Systems, Saso, 2015-october, 2015, pp. and of the 2012 Rutgers School of Engineering Distinguished 101–110. Young Alumnus Medal of Excellence. She is a coauthor of the [10] A. Pantoja and N. Quijano, “Distributed optimization book Wireless Networks: Multiuser Detection in Cross-Layer using population dynamics with a local replicator,” Design (Springer, NY). Her research interests are focused on Proceedings of the IEEE Conference on Decision and applications of game theory, evolutionary games and machine Control, vol. 3790-3795, 2012, pp. 3790–3795. learning for resource management and optimization of complex distributed networks. Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Quantitative Finance arXiv (Cornell University)

Dynamic Influence on Replicator Evolution for the Propagation of Competing Technologies

Bolluyt, Elijah D.; Comaniciu, Cristina
Quantitative Finance , Volume 2019 (1911) – Nov 8, 2019
Download PDF
6 pages

Loading next page...
 
/lp/arxiv-cornell-university/dynamic-influence-on-replicator-evolution-for-the-propagation-of-U1DGkN9VYg

References (14)

ISSN
1089-778X
eISSN
ARCH-3346
DOI
10.1109/TEVC.2018.2881973
Publisher site
See Article on Publisher Site

Abstract

This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 Dynamic Influence on Replicator Evolution for the Propagation of Competing Technologies Elijah D. Bolluyt, Comaniciu C., Member IEEE factors on individualization of strategies [3][10]; however, Abstract— This work introduces a novel modified Replicator while including some environmental dynamism, these Dynamics model, which includes external influences on the approaches still do not model an environment in which population. This framework models a realistic market into which external parties actively influence the market over time, while companies, the external dynamic influences, invest resources in the individuals in the population remain homogeneous. order to bolster their product’s standing and increase their Additionally, previous applications of the replicator market share. The dynamic influences change in each time step framework to realistic scenarios have not focused on dynamic of the game, and directly modify the payoff matrix of the population’s interactions. The model can learn from real data modification over time to the payoff matrix [9][11][12][13]. In how each influence affects the market, and can be used to addition, to the best of our knowledge, no work exists in simulate and predict the outcome of a real system. We integrating real data into such a model in order to make specifically analyze how a new technology can compete and predictions on a real world system. These characteristics are attempt to unseat an entrenched technology as the market leader. necessary to create a model that can accurately emulate and We establish a relationship between the external influences and predict a realistic marketplace. the population payoff matrix and show how the system can be implemented to predict outcomes in a real market by simulating This work introduces a general method to realistically the rise of the Android mobile operating system over its primary model the dynamic influence of external factors on an competition, the iPhone, from 2009-2017. unmodified population, and investigates the actions necessary for such an outside influence to achieve specific goals. We Index Terms— Evolutionary game theory, Product markets, then apply this method to specifically address the case of a Replicator dynamics market that is initially dominated by an entrenched technology, describing the conditions necessary for a smaller, I. INTRODUCTION emerging technology to compete and, eventually, overtake the VERY real environment is in a state of perpetual flux, market leader. Our model not only predicts market conditions especially those that share a close relationship with based on input information, but it also provides the user with technology. As technologies evolve in a free market, information about the intrinsic properties of the marketplace. competition inevitably arises; sometimes, one company Because our approach emphasizes realism and applications dominates, and other companies expend resources to make for modeling real markets, we conduct a simulation of the real inroads. No matter the specifics, competition leads to each case of the world’s smartphone market from 2009-2017 to individual in the market choosing which technology to use prove the efficacy of our model. This range captures the rise based on the relative value of each to the user. of the Android smartphone operating system (OS) and its Replicator dynamics provides a useful framework in which ascension to dominance over the Apple iPhone, which had to analyze competitive markets, as it models a population of hitherto led the market [4]. identical individuals who choose a strategy based on its projected payoff, or the value to the user [1]. However, the II. MODEL original replicator equation’s payoff is a function only of the A. General Replicator Dynamics Model portion of the population which employs each strategy; the strategy with the larger user base dominates [1]. In a free and Replicator dynamics models the evolution of a population of active market, however, each rational company will invest identical individuals who employ one of n strategies and assets to increase its product’s adoption rate, and thus its receive a payoff from each interaction with other individuals based only on each individual’s choice of strategy [1]. A profits. player employing strategy i who interacts with one employing Previous extensions of the replicator dynamic model have a strategy j receives a payoff 𝐴𝐴 ; the set of payoffs can be included feedback loops between the environment and the expressed as a matrix: strategy payoffs [2], as well as the effects of localization 𝐴𝐴 𝐴𝐴 ⋯ 𝐴𝐴 11 12 1𝑛𝑛 Manuscript submitted February 12, 2018; revised June 24, 2018 and 𝐴𝐴 𝐴𝐴 ⋯ 𝐴𝐴 21 22 2𝑛𝑛 September 10, 2018. 𝐴𝐴 = � � (1) ⋮ ⋮ ⋱ ⋮ E. Bolluyt and C. Comaniciu are with the Department of Electrical and Computer Engineering, Stevens Institute of Technology, Hoboken, NJ 07030 𝐴𝐴 𝐴𝐴 ⋯ 𝐴𝐴 𝑛𝑛1 𝑛𝑛2 𝑛𝑛𝑛𝑛 USA (e-mail: ebolluyt@stevens.edu; ccomanic@stevens.edu). Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. 𝑖𝑖𝑖𝑖 This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 A fraction 𝑥𝑥 ∈ [0,1] of the population employs the strategy 𝑥𝑥 ̇ > 0 → 𝐴𝐴 𝑥𝑥 +𝐴𝐴 𝑥𝑥 > 𝐴𝐴 𝑥𝑥 +𝐴𝐴 𝑥𝑥 (6) 𝑖𝑖 1 11 1 12 2 21 1 22 2 i. The replicator equation expresses the growth rate of this portion of the population: When combined with (5), we see that growth of 𝑥𝑥 requires: 𝑥𝑥 >�𝑥𝑥 (7) 1 1 𝑇𝑇 𝑥𝑥 ̇ = 𝑥𝑥 ((𝐴𝐴 𝑥𝑥 ) −𝑥𝑥 𝐴𝐴 𝑥𝑥 ) (2) 𝚤𝚤 𝑖𝑖 𝑖𝑖 This condition implies that the system moves away from this mixed equilibrium, not towards it, making this point an A strategy’s usage only increases if its payoff is higher than unstable equilibrium point; it also tells us that the payoff the average payoff in the population. Thus, the expected matrix A must favor a low value mixed equilibrium in order payoff of each strategy directly modifies the individuals’ for 𝑥𝑥 to continue to grow. choices of which strategy to employ. Condition (7), combined with (2) and (6), allow us to The growth expressed by the replicator equation (2) analyze the effects and desirable traits of each of the four captures the change in the fractions of the population that payoff elements from the standpoint of strategy 1’s proponent, employ each strategy over time, showing how a higher payoff the external influence working in favor of strategy 1. causes a strategy to gain prominence in the population. should be increased according to (2). To attain growth, 𝐴𝐴 Note that for the population’s strategy distribution to change This compensates for the small initial value of 𝑥𝑥 . As 𝑥𝑥 1 1 significantly, the elements of A must be very different. It can approaches dominance, 𝐴𝐴 can be allowed to decrease as be seen from (2) that if all elements of A are approximately lower growth is expected. 𝐴𝐴 should be kept large in the equal, then the payoff of each strategy will be very close to the beginning to allow for the dominance of strategy 1, but its average, and thus no strategy’s growth rate will be significant. value becomes less important as 𝑥𝑥 increases. 𝐴𝐴 should also 1 21 B. Two Strategy Replicator Dynamics Model be kept large enough to allow for dominance of 𝑥𝑥 , but it should decrease as 𝑥𝑥 increases as this will allow the growth To model a technology market with one entrenched, rate of 𝑥𝑥 to increase. dominant technology and one new technology, we formulate a Note that 𝐴𝐴 is not under the control of an influence of population with two strategies available to each individual, 22 strategy 1. A technology’s interactions with itself should be where each strategy represents the usage of one of the two controlled only by the company that owns it. We can expect, technologies. The payoff matrix A becomes: however, that in a population where 𝑥𝑥 is increasing, 𝐴𝐴 is 1 22 likely to be small in magnitude. 𝐴𝐴 𝐴𝐴 11 12 𝐴𝐴 = � � (3) 𝐴𝐴 𝐴𝐴 21 22 C. Dynamic Influence Extension We now formulate an extension to the payoff matrix that 1) Equilibrium Conditions allows such modifications, as outlined above, to be made by The two strategy replicator dynamics model has three an external influence on the market. We construct a column equilibrium points at which the growth of each strategy is vector where each element 𝑦𝑦 ,𝑖𝑖 ∈ [1,𝑛𝑛 ] represents an zero: 𝑖𝑖 𝑦𝑦 𝑥𝑥 ̇ = 𝑥𝑥 ̇ = 0 (4) influencing factor. These factors are inputs that affect the 1 2 market and can be modified by the policies of companies in Two are trivial, wherein one strategy grows to engulf the the market, such as product pricing or investments (which entire market and the other is driven to extinction. The third directly impact the quality of the product). The number of the equilibrium point can be found by using (2) and (4): inputs 𝑛𝑛 needed to accurately characterize the markets will 𝑦𝑦 depend entirely on the properties of the markets under 𝐴𝐴 −𝐴𝐴 22 12 �𝑥𝑥 = (5) 1 investigation, and has no direct relation to 𝑛𝑛 . The system will (𝐴𝐴 +𝐴𝐴 )−(𝐴𝐴 +𝐴𝐴 ) 11 22 21 12 be able to model the market evolution only if the major relevant inputs that characterize that particular market are It is important to note that the equilibrium conditions reached are dependent not only on the system’s parameters, but also on identified. the initial conditions; if one strategy starts with a market share Considering n influencing input factors, each element of of zero, then it never grows and the system stays constant. the payoff matrix becomes a linear combination of these factors. The linear coefficients are expressed as a newly 2) Impact of Payoff Matrix Element Values on System defined 𝑛𝑛 ×𝑛𝑛 matrix 𝛼𝛼 . To illustrate the influence of the 𝑦𝑦 Evolution input parameters on the elements of matrix A, the payoff The elements of the payoff matrix directly influence the matrix A can be re-indexed with a single index 𝑘𝑘 ∈ [1,𝑛𝑛 ], outcome of the market. Thus, the behavior of the system can ( ) traversing row by row, where 𝑘𝑘 = 𝑛𝑛 𝑖𝑖 − 1 +𝑗𝑗 and i,j are the be changed by modifying the elements of the payoff matrix in original indices 𝑖𝑖 ,𝑗𝑗 ∈ [1,𝑛𝑛 ]. Consequently, each element of specific ways. An external party looking to influence the the payoff matrix can be expressed as a linear combination of market should attempt to make its strategy (i.e. technology) the input factors: constantly grow and finally reach equilibrium at a dominant 𝐴𝐴 = 𝛼𝛼 𝑦𝑦 (8) position. 𝑘𝑘 𝑘𝑘 where 𝛼𝛼 is the kth row of the coefficient matrix 𝛼𝛼 . In order to grow, a strategy’s payoff must be greater than 𝑘𝑘 This representation allows the external influences to have a the average; (2) indicates that this condition is necessary for direct impact on the outcome of the market. As stated above, the growth rate to be positive. This constraint leads to a few constraints must be placed on this matrix 𝛼𝛼 to maintain condition (6) Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 realism; specifically, any factor controlled by company i must only affect terms involving strategy i. This modification makes A entirely dynamic, able to take on different values during each iteration of the simulation. This allows the model to take into account the influence of external parties which change their strategies based on the market. An external party can evaluate how its input affects the payoff matrix and thus the market, and react by changing its input values accordingly. These factors may include investments a company makes into its product, pricing of its product, or any other values that have a direct impact on the quality and availability of its technology. The value of 𝛼𝛼 is a constant, and it represents an intrinsic property of the market’s interactions with the input factors. For any particular market and input set y, 𝛼𝛼 must be learned using real data. The value of 𝛼𝛼 gives the user insight into the market, showing the influence of each input on the utility of Fig 2. System outputs. The inflection point in the market and the values in each product. This stands in contrast to other learning the validation points are accurately predicted by the system. methods, in which the learned parameters give no insight into the simulated system. necessarily reach this equilibrium, however, because the In our simulation, we employ a simple exhaustive search system changes over time. The system converges to a final and evaluate candidate results via a least mean squares error equilibrium only when the inputs y, and thus the payoff metric, while withholding 20% of the available data in matrix, converge to a constant value. accordance with common practice [5]. When training the system, learning the parameters of A, this last 20% of data was III. SIMULATION withheld from the set; this tests the model’s predictive A. Setup and Data capability. A mean squared error metric was used as it is To validate the accuracy of the proposed model for practical computationally simple and accurately represents the validity applications, a simulation was conducted on data from the of the result. global smartphone market from Quarter 1 (Q1) 2009 to Q1 𝑡𝑡 2 𝑓𝑓 𝐽𝐽 = �𝑥𝑥 (𝑡𝑡 )−𝑥𝑥 (𝑡𝑡 )� (9) 1 1 𝑡𝑡=1 2017. This range was chosen as it captures the rise of the 𝑡𝑡 𝑓𝑓 Google Android platform to dominance over the Apple iPhone Where 𝑥𝑥 (𝑡𝑡 ) is the prediction at time 𝑡𝑡 and 𝑥𝑥 (𝑡𝑡 ) is the real in market share. In order to simulate a two-strategy system, data at time 𝑡𝑡 . the smartphone market is simplified to just Android and iOS; Due to the time dynamic nature of these modifications, the while other platforms existed, these two grew to dominate the system no longer has a single guaranteed equilibrium value to market as the modern wave of touchscreen smartphones which it will converge. At each time step, the system has a developed [4]. target equilibrium to which it would converge if the system The system takes four input factors 𝑦𝑦 : average pricing for 𝑖𝑖 inputs y were to stay constant. The system does not Fig 1. System inputs. These inputs contribute directly to the growth or Fig 3. Payoff Matrix Coefficients. These four computed factors determine shrinkage of each technology in the market. Note the difference in the market behavior as discussed in Section II, and their values are a direct magnitudes between the investment factors, as well as the opposing trend result of the system inputs. Note the small magnitude of A as well as the directions of the products’ pricing. trend directions of A , A , and A . 11 12 21 Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 TABLE I OPTIMIZED VALUES OF 𝛼𝛼 𝒚𝒚 𝒚𝒚 𝒚𝒚 𝒚𝒚 𝟏𝟏 𝟐𝟐 𝟑𝟑 𝟒𝟒 𝑨𝑨 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝛼𝛼 11 12 13 14 𝑨𝑨 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝟐𝟐𝟏𝟏 21 22 23 24 Simulation 𝑨𝑨 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝟏𝟏𝟐𝟐 31 32 33 34 𝑨𝑨 𝛼𝛼 𝛼𝛼 𝛼𝛼 𝛼𝛼 41 42 43 44 Real Data 4 0 -1 0 0 3 1 3 3 0 3 1 0 4 0 -1 Constant Market 4 0 2 0 (iOS dominant) 0 3 -3 4 3 0 4 -3 0 4 0 2 The value structure sho ws to which input 𝑦𝑦 and payoff matrix element 𝑖𝑖 𝐴𝐴 each element of 𝛼𝛼 belongs, as each is a linear coefficient for one input in one payoff element. 𝑦𝑦 : Android Investment 𝑦𝑦 : Android Price 1 3 𝑦𝑦 : iOS Investment 𝑦𝑦 : iOS Price Fig 4. System outputs in simulation with constant inputs, using 𝛼𝛼 optimized 2 4 with real data. This shows that the strategy Android employed to gain dominance in the market had a great effect on increasing its final market C. Conditions on 𝛼𝛼 for Optimization share. A is normalized to keep all elements 𝐴𝐴 ∈ [0,1], in Android and iOS smartphones sold each quarter [8], and accordance with the standard replicator equation conditions quarterly research investments in mobile from Apple [6] and (1). Thus the magnitude of 𝛼𝛼 does not matter, as the system’s Samsung [7], the largest Android phone manufacturer. These behavior is determined by the relative size of its elements. factors were chosen as they are determined by the device The grid search was conducted over an interval [−𝑟𝑟 ,𝑟𝑟 ] by manufacturers’ direct influence and have a direct impact on integer steps; the search domain was optimized as a hyper the devices’ market performance. The values of these factors parameter of the system over multiple trials, and a value of over the simulation period are shown in Fig. 1; these values 𝑟𝑟 = 4 was chosen for this simulation. The complexity of the were obtained directly from real data [6][7][8]. The data was 𝑛𝑛 ∗𝑛𝑛 /2 𝑦𝑦 search increases on the order of (2𝑟𝑟 + 1) , or (2𝑟𝑟 + 1) sampled quarterly, to match the available frequency of in this particular simulation; this makes increasing the search investment data from the companies’ quarterly reports. This interval very computationally expensive. The value of 𝑟𝑟 gave a total of 33 time samples in the simulation period. Even determines the resolution of the resulting 𝛼𝛼 , and must be with this relatively sparse data, the model remained effective; chosen by trial and error. For our simulation we started with results are discussed further in Section III-D. 𝑟𝑟 = 1 and increased 𝑟𝑟 until the error between the simulated The system is initialized at the market shares determined by −5 curve and the real market evolution was less than 4 × 10 . inputting the first real data point into the system; this is vital to Since the search is conducted by integer steps and the obtaining a result that mirrors the reality of the market. The resulting A is normalized, an interval of [−𝑟𝑟 ,𝑟𝑟 ] is equivalent initial conditions are central to the determination of the to searching on an interval of [−1,1] with a step size of 1/𝑟𝑟. A system’s behavior, as mentioned in Section II-B, so they must higher value for 𝑟𝑟 is desirable for better resolution of the be based on the real data. search, however a high value for r induces a higher As discussed in IIC, a subset consisting of approximately computational complexity. The complexity increases with 20% of the latest data, or 7 time samples, is withheld from the both the number of shares in the market, 𝑛𝑛 , and the number of training set to serve as a validation set in order to determine input factors 𝑛𝑛 . This model is well suited to technological the system’s predictive capability. 𝑦𝑦 markets, where typically only a few companies have the B. Implementation capability to compete effectively. Technological markets also often have a high barrier to entry, making the study of the A grid search is employed to optimize 𝛼𝛼 . For each candidate confict between a dominant product and a new competitor value, the complete evolution of the system over the training especially relevant. interval is computed. For each time step in the evolution, the Imposing practical, common sense constraints, the effect of matrix A is computed as defined in (8) and normalized, and the growth rates of the two technologies (Android and iOS) the companies’ influences on their respective products must be identical; a change in an input parameter for product 1 must are computed (2). The market values at all time-steps are computed using the A and growth rate obtained for the cause a variation in the payoff coefficients symmetric to the result of the same change in the same input parameter for candidate 𝛼𝛼 value, and the error J between this prediction and product 2. If parameters 𝑦𝑦 and 𝑦𝑦 are the same input for the the real data is computed. The value of 𝛼𝛼 with the lowest error 1 2 two respective products, then the rates of change of A’s over the training interval is chosen. This 𝛼𝛼 is used for elements must be constrained by validation in section D. Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. 𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 𝟐𝟐𝟐𝟐 𝟏𝟏𝟏𝟏 This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 𝜕𝜕𝐴𝐴 𝜕𝜕𝐴𝐴 E. Analysis and Significance of 𝛼𝛼 { } = ;𝑖𝑖 ,𝑗𝑗 = 1,2 (10) 𝜕𝜕𝑦𝑦 𝜕𝜕𝑦𝑦 𝑖𝑖 𝑖𝑖 The value of 𝛼𝛼 obtained from the simulation using real data Since these partial derivatives are linear, this can be is shown in Table 1. The values of its individual elements accomplished by constraining the values of the elements of 𝛼𝛼 show the effect of each input on the payoff matrix, and thus which correspond to these enforced relations to be equal the market. As can be expected, higher investment of each because the elements of 𝛼𝛼 are the rates of change of A’s company in its own product positively effects its own elements with respect to the input parameters y. These and α ), and higher pricing of each product’s utility (α 11 42 constraints ensure that the system remains symmetric. If we product negatively effects its utility (α and α ). Similarly, 13 44 linearly index A with a single variable 𝑙𝑙 , traversing row by row increased research spending also increases the utility that the 𝜕𝜕𝐴𝐴 𝑙𝑙 as stated in Section IIC, corresponds to 𝛼𝛼 as each row of product’s users obtain from interacting with the competitor 𝑙𝑙 𝑘𝑘 𝜕𝜕𝑦𝑦 𝑘𝑘 technology (α and α ), indicating efforts to increase cross 𝛼𝛼 corresponds to one element of A, as expressed in (8). 22 31 compatibility to increase market uptake. Note that any influence on technology i should not exert an A second simulation was run, in which the optimized value influence on 𝐴𝐴 , as discussed in Section IIB. Thus the 𝑖𝑖𝑖𝑖 of α was used to predict a market outcome where the elements of 𝛼𝛼 corresponding to these factors are zero. In a companies kept a constant level of investment and pricing, two strategy population, there will be one such factor for each equal to that in the first time step. This was done to assess the input as each input corresponds exclusively to one technology. effectiveness of the strategies employed by each company, and D. Analysis determine what effect the feedback from the market had on the outcome. The results of this simulation can be seen in Fig. 4; The simulated and real data can be seen in Fig. 2. The Android still becomes dominant, but its market share saturates model was trained using the training set, and then the system at a lower value. This would seem to indicate that Android’s was allowed to evolve on its own, employing its learned value strategy of decreasing price and increasing investment payed of α and the real market input data (prices and investments) off significantly. for the duration corresponding to the validation set. Its Finally, a third simulation was conducted in which the computed market share values had a mean squared error −5 market shares were held constant. This was done to observe of 3.0 × 10 with respect to the real data for this period. what type of market would have allowed iOS to remain This reflects the ability of the system to predict future data dominant, given the real input data. The value of α optimized with high accuracy. While this simulation approached an for this hypothetical situation is also shown in Table 1. equilibrium state, this is due to the inputs remaining fairly Investments play a similar role in increasing a product’s constant near the end of the simulation period; such a utility, but higher prices now have a large positive effect on convergence is not guaranteed in this model, as discussed in the utility of each product (α , α ). Such an unrealistic Section II-C. 13 44 property makes sense because iOS devices have a much higher The values of A’s elements (Fig. 3) in the simulation price compared to Android devices throughout the simulation conformed well to expected behaviors from Section IIB. 𝐴𝐴 period. If consumers wanted to pay a higher price, then this increases while technology 1 (Android) holds a small market would have improved iOS’s market share size rather than share, and starts to decrease after the inflection point in the shrinking it. market. 𝐴𝐴 remains large with some fluctuation throughout Analyzing these simulations shows that the properties of α the simulation. 𝐴𝐴 starts large to allow for initial growth, reflect real properties of the market. This direct relation is the then decreases as Android reaches a dominant market position. core advantage of this model over other learning based 𝐴𝐴 is small in magnitude in comparison to the other methods: the learned parameters give direct insight into the coefficients; this is to be expected in a market where market reaction relative to the inputs. This could potentially technology 1 is growing its market share. However, 𝐴𝐴 did allow a market player to not only simulate strategies and increase over the simulation period, reflecting Apple’s observe the outcome, but also to understand how the increased investment in iOS and their pricing decreases near individual inputs under its control could modify the market the end of the observed period, likely an effort to grow iOS’s share evolution. The players could then devise a strategy that market share. achieves the desired outcome based on this knowledge of the The system’s accuracy to both the real market values and market. the expected internal behavior indicate its efficacy in modelling the marketplace. The small input data size of 33 IV. CONCLUSION time steps coupled with these results indicate that the system is able to perform with much less data than would be required This work extends the Replicator Dynamics framework to for many learning based models. Its replicator foundation analyze a system with external influencing factors. It modifies specifically models strategy fitness as a function of the payoff matrix of the core system to adapt to external population, an important feature when dealing with factors, and learns parameters to model that dependence. This technologies that require a user base to stay viable. Versus a enables the system to contend with external factors that general learning method, the model only has to learn the influence an otherwise symmetric population, an ability not specific nature of a particular market rather than learning the present in previous replicator formulations. This modified nature of markets in general; this allows it to perform well framework was applied to the rise of the Android platform to with less data, as less information has to be learned. mobile OS market dominance between 2009 and 2017; the proposed system accurately modeled the evolution of each Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org. 𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TEVC.2018.2881973 technology’s market share for several fiscal quarters, [11] J. Huang, J. Wang, H. Zhang, and N. Wang, “2017 employing parameters obtained via optimization over a Ieee 2nd Advanced Information Technology, training set. The simulation also produced intermediate Electronic and Automation Control Conference parameter behavior which closely corresponded with (Iaeac).” Network Defense Strategy Selection Based theoretical predictions for a scenario such as the rise of on Best-Response Dynamic Evolutionary Game Android over iOS, starting from a small initial market share. Model, IEEE, 2017, pp. 2611–2615. This simulation result indicates that the system can be [12] S. D’Oro, L. Galluccio, S. Palazzo, and G. Schembra, employed to accurately model and predict the evolution of “A Game Theoretic Approach for Distributed markets in real-world scenarios. This model could aid in Resource Allocation and Orchestration of predicting the outcome of variations in strategy of companies Softwarized Networks.” IEEE Journal on Selected in a marketplace, making it a useful tool to create new Areas in Communications, vol. 35, no. 3, 2017, pp. strategies which can produce a desired outcome. 721–735. [13] M. A. Shattal, A. Wisnieswska, A. Al-Fugaha, B. Khan, K. Dombrowski, “Evolutionary Game Theory REFERENCES Perspective on Dynamic Spectrum Access Etiquette.” [1] R. Cressman, “Symmetric Normal Form Games,” in IEEE Access, (2017 08 22), 2017. Evolutionary Dynamics and Extensive Form Games, Cambridge, MA, USA: The MIT Press, 2003, ch. 2, Elijah D. Bolluyt received the B.S. degree sec. 1, pp.19-22. in engineering physics and the M.E. degree [2] J.S. Weitz, C. Eksin, K. Paarporn, S. P. Brown, and in electrical engineering from Stevens W. C. Ratcliff, “An oscillating tragedy of the Institute of Technology, Hoboken, NJ, in commons in replicator dynamics with game- 2017. He is currently pursuing the Ph.D. environment feedback,” PNAS, vol. 113, no. 47, pp. degree in computer engineering at Stevens E7518-E7525, Nov. 2016. [Online]. DOI: Institute of Technology, Hoboken, NJ. 10.1073/pnas.1604096113. During the summer of 2015, he was a [3] F. Pieper and S. Mostaghim, "Influence of dynamic Scholars summer researcher with the Department of Physics environments on agent strategies," 2016 IEEE and Engineering Physics at Stevens Institute of Technology. Symposium Series on Computational Intelligence During the summer of 2016, he was an Engineering Intern at (SSCI), Athens, 2016, pp. 1-8. doi: 10.1109/SSCI. Kulite Semiconductor, Inc. in Leonia, NJ. His current 2016.7850159 research interests focus on the development of computational [4] “Mobile Operating System Market Share frameworks, employing game theory and machine learning Worldwide,” StatCounter Global Stats. Accessed techniques. September 15, 2017. [Online]. Available: Mr. Bolluyt is a recipient of the Provost’s Doctoral http://gs.statcounter.com/os-market-share/mobile/ Fellowship (Stevens Institute of Technology). worldwide/ [5] T. M. Mitchell, “Decision Tree Learning,” in Cristina Comaniciu received the M.S. Machine Learning. New York, NY, USA: McGraw- degree in electronics from the Polytechnic Hill, 1997, ch. 3, sec. 7.1, pp.68-70. University of Bucharest in 1993, and the [6] “Earnings Releases,” Apple Inc., Investor Relations. Ph.D. degree in electrical and computer Accessed September 15, 2017. [Online]. Available: engineering from Rutgers University in http://investor.apple.com/financials.cfm [7] “Earnings Releases,” Samsung Global, Investor From 2002 to 2003 she was a Relations. Accessed September 15, 2017. [Online]. postdoctoral fellow with the Department Available: http://www.samsung.com/global/ir/ of Electrical Engineering, Princeton University. Since August financial-information/earnings-release/ 2003, she is with Stevens Institute of Technology, Department [8] F. Richter, “The Smartphone Price Gap,” Statista, of Electrical and Computer Engineering, where she is now an June 2, 2016. [Online]. Available: Associate Professor and serves as Associate Department Chair https://www.statista.com/chart/4954/smartphone- for Graduate Studies. In Fall 2011 she was a visiting faculty average-selling-prices/ fellow with the Department of Electrical Engineering, [9] K. Coninx and T. Holvoet, “Darwin in Smart Power Princeton University. She served as an associate editor for the Grids – Evolutionary Game Theory for Analyzing IEEE COMMUNICATION LETTERS (2007-2011). Self-Organization in Demand-Side Aggregation,” Professor Comaniciu is a recipient of the 2007 IEEE International Conference on Self-Adaptive and Self- Marconi Best Paper Prize Award in Wireless Communications Organizing Systems, Saso, 2015-october, 2015, pp. and of the 2012 Rutgers School of Engineering Distinguished 101–110. Young Alumnus Medal of Excellence. She is a coauthor of the [10] A. Pantoja and N. Quijano, “Distributed optimization book Wireless Networks: Multiuser Detection in Cross-Layer using population dynamics with a local replicator,” Design (Springer, NY). Her research interests are focused on Proceedings of the IEEE Conference on Decision and applications of game theory, evolutionary games and machine Control, vol. 3790-3795, 2012, pp. 3790–3795. learning for resource management and optimization of complex distributed networks. Copyright (c) 2018 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.

Journal

Quantitative Finance – arXiv (Cornell University)

Published: Nov 8, 2019

There are no references for this article.