Learning, Games, and Networks Seminar

The Learning, Games, and Networks Seminar was organised by Daniel Quigley and myself during the 2016/7 academic year. On this page, you can find information about past talks. In the 2017/8 academic year, this seminar will be organised by Daniel Quigley, and David Ronayne. For information of upcoming talks, please go to this webpage.

Past talks during the 2016/7 academic year::

October 11st, 2016, at 16:30 (Week 1 of Michaelmas):

Speaker: Edoardo Gallo, University of Cambridge
Venue: Staircase L, Conference Room (Upstairs)
Title: Financial contagion in networks: An experiment


The 2008 global financial crisis highlighted the crucial role that the network architecture of the financial system plays in determining systemic contagion. In the aftermath of the crisis, the Bank of England argued that “the financial network should be structured so as to reduce the chances of future systemic collapse” and “better information on connections between firms in the financial network [is crucial to] building a more resilient financial system”.

This paper investigates experimentally the roles of network and information structures on financial contagion, price formation, and the behaviour of traders. Participants have heterogeneous valuations for assets and they are assigned to a position in a network of liabilities that leaves them exposed to counterparty risk. One participant is hit by a shock whose size is common knowledge. Participants can trade assets in a double auction market and they face a trade-off: buy to earn a long-term return from the assets vs. sell to raise liquidity to cushion the potential spillovers from the shock.

Network structure has a significant impact on the resilience of the system to shocks. Financial contagion and individual bankruptcy are much more likely in core-periphery compared to circle networks. In core-periphery networks, the traders perceive this heightened risk leading to a collapse in prices and a market freeze where everyone is trying to sell assets. In contrast, in circle networks the market functions normally. Whether market participants have information about the location of the shock in the network, however, has no substantial effect on financial contagion, individual bankruptcy, the evolution of prices, or traders’ bidding behaviour.

October 25th, 2016, at 16:30 (Week 3 of Michaelmas):

Speaker: Juan Block, University of Cambridge
Venue: Staircase L, Large Lecture Room
Title: Learning Dynamics Based on Social Comparisons

Abstract: We study models of learning in games where agents with limited memory use social information to decide when and how to change their play. When agents only observe the aggregate distribution of payoffs and only recall information from the last period, we show that aggregate play comes close to Nash equilibrium behavior for (generic) games, and that pure equilibria are generally more stable than mixed equilibria. When agents observe not only the payoff distribution of other agents but also the actions that led to those payoffs, and can remember this for some time, the length of memory plays a key role. When agents’ memory is short, aggregate play may not come close to Nash equilibrium, but it does so if the game satisfies a acyclicity condition. When agents have sufficiently long memory their behavior comes close to Nash equilibrium for generic games. However, unlike in the model where social information is solely about how well other agents are doing, mixed equilibria can be favored over pure ones.

November 8th, 2016, at 16:30 (Week 5 of Michaelmas):

Speaker: Ina Taneva, University of Edinburgh
Venue: Staircase L, Large Lecture Room
Title: Information Design: The Epistemic Approach

Abstract: Information design studies how to disclose information to a group of interacting agents in order to influence their behavior. In this paper, we introduce a belief-based approach to the problem, viewing it as belief manipulation rather than as information disclosure. We characterize and then exploit the equivalence between information structures and distributions over belief hierarchies. Our main result is a representation theorem that poses the design problem as a choice of an optimal distribution over a special family of belief-hierarchy distributions—the minimal consistent ones—subject to Bayes plausibility. A two-step decomposition of the theorem follows, leading to a concave-envelope representation of optimality that subsumes Kamenica and Gentzkow (2011)’s single-agent result. We apply our representation theorem to a managerial problem, where we study Bayes Nash information design, and to a classic investment game, where we study information design under bounded depths of reasoning.

November 22nd, 2016, at 16:30 (Week 7 of Michaelmas):

Speaker: Heinrich Nax, ETH Zurich
Venue: Staircase L, Conference Room (Upstairs)
Title: Market behavior when more information becomes available over time (joint with Peiran Jiao)

Abstract: We conducted a controlled laboratory experiment to understand how arrival of different types of information changes the decision-making process of subjects involved in a repeated game. We considered the case of Cournot oligopolies and study convergence properties of the resulting price evolution as more information is revealed. Our findings suggest that low-information decision-making processes continue to dominate, provided convergence to Nash-equilibrium prices has already progressed substantially in the preceding low-information phases. Collusion is not observed. Crucially, explicit information about relative profits may trigger reciprocity and imitation dynamics that lead from Nash equilibrium toward the zero-profit Walrasian outcome.

January 17, 2016, at 16:30 (Week 1 of Hilary):

Speaker: Sam Jindani, University of Oxford
Venue: Staircase L, Conference Room (Upstairs)
Title: Are social norms equilibria?

Abstract: Why did footbinding in China endure for a millennium and then suddenly die out? Why did duelling disappear abruptly in the United Kingdom but slowly decline in France? I present a simple model of social norms that explains these phenomena. The model predicts that the evolution of norms is characterised by tipping, whereby norms can shift suddenly due to shocks, and by a ratchet effect, whereby changes in parameters can cause norms to decline gradually. I show that the model can be supported by an equilibrium of a repeated game, with no special assumptions about preferences.

January 31, 2016, at 16:30 (Week 3 of Hilary):

Speaker: Marco Pangallo, Institute for New Economic Thinking at the Oxford Martin School; Mathematical Institute, University of Oxford
Venue: Staircase L, Conference Room (Upstairs)
Title: Does learning converge in generic games? (joint with J. D. Farmer, T. Galla, T. Heinrich, J. Sanders)

Abstract: In game theory, learning has often been proposed as a convincing method to achieve coordination on an equilibrium. But does learning converge, and to what? Most of the existing literature focuses on specific classes of games and concludes that convergence is at least prevalent. However, some recent work showed that learning in ‘typical’ games with many moves follows highly chaotic dynamics.

In order to shed light on the mechanisms behind (non-)convergence, we start investigating the drivers of instability in the simplest possible non-trivial setting, that is generic 2-person, 2-strategy normal form games. We use Experience-Weighted Attraction (EWA), which encompasses most extensively studied learning algorithms and has been shown to be in accord with experimental data. We exhaustively characterize the parameter space of EWA learning, for any payoff matrix, and we obtain an emerging taxonomy of learning dynamics that depends on some structures in the payoff matrix and on the ‘speed’ of learning. In games with a unique mixed strategy equilibrium the players follow the cycle of best responses and never converge to the Nash Equilibrium: we rather observe limit cycles or low-dimensional chaos.

We then characterize the cyclic structure of games with many moves as a sufficient statistics for the convergence of learning: we obtain convergence in coordination, potential, supermodular and generally acyclic games, but such games become more and more rare as the number of moves increases (a fortiori if the payoffs are negatively correlated and with more than two players). In all other games, the learning dynamics may follow limit cycles or end up in chaotic attractors, depending on the details of the payoffs and on the degree of sophistication of the learning algorithm. If the aforementioned types of games are not representative of real economic situations, strategic interactions would be governed by learning in an ever-changing environment, rather than by rational and fully-informed equilibrium thinking.

February 7, 2016, at 16:30 (Week 4 of Hilary):

Speaker: Dimitri Migrow, University of Manchester
Venue: Staircase L, Conference Room (Upstairs)
Title: Designing Communication Hierarchies to Elicit Information

Abstract: A central problem in organizations is that much of the information relevant for decision making is dispersed among employees who are biased and may lack the incentives to communicate their information to the management. This paper studies how a manager can elicit employees’ information by designing a hierarchical communication network. The manager decides who communicates with whom, and in which order, where communication takes the form of “cheap talk” (Crawford and Sobel, 1982). I show that the optimal network is shaped by two competing forces: an intermediation force that calls for grouping employees together and an uncertainty force that favors separating them. The manager optimally divides employees into groups of similar bias. Under simple conditions, the optimal network features a single intermediary who communicates directly to the manager.

February 14, 2016, at 16:30 (Week 5 of Hilary):

Speaker: David Sturrock, Institute for Fiscal Studies
Venue: Staircase L, Conference Room (Upstairs)
Title: Commitments and Partnerships (joint with Yuval Heller)

Abstract: We explore a 2-player partnership game where, before choosing a level of effort to exert on a joint project, each player sends a message to their partner, interpreted as a commitment about future effort. We allow that a convex cost of ‘reneging’ – the distance between a player’s commitment and their actual effort – may enter the players’ subjective utility functions, but this is not part of their material payoffs. We study the endogenous determination of this ‘reneging cost’ in an evolutionary setting. When each player may observe their partner’s reneging cost with a high probability, there is a unique stable state in which all players have the same ‘intermediate’ reneging cost, and the equilibrium effort is a second-best outcome.

February 28, 2016, at 16:30 (Week 7 of Hilary):

Speaker: Bary Pradelski, ETH Zurich
Venue: Staircase L, Conference Room (Upstairs)
Title: Social influence and opinion polarization on the Internet: a field experiment (joint with Bernhard Clemm von Hohenberg, Michael Mäs)

Abstract: Does social influence on the Internet intensify opinion polarization? We conducted a field experiment (N = 3,632) on a news website where readers can indicate their opinion after seeing previous users’ opinions. Calibrating an econometric model that captures prominent micro-models of social influence, we find that opinions shift towards the average view of earlier readers. Disagreement with others’ opinions does not reduce influence or motivate shifts away. With this calibrated micro-model we predict the macro-level opinion dynamics resulting from social influence. As predicted, we empirically observe that influence does not affect the average of users’ opinions and reduces opinion polarization.

May 9, 2017, at 14:00 (Week 3 of Trinity):

Speaker: Matthew Elliott, Cambridge
Venue: Staircase L, Conference Room (Upstairs)
Title: Decentralized Bargaining: Efficiency and the Core (Joint with Francesco Nava)

Abstract: This paper studies market clearing in matching markets. The model is non-cooperative, fully decentralized, and in Markov strategies. Workers and firms bargain with each other to determine who will be matched to whom and at what terms of trade. Once a worker- firm pair reach agreement they exit the market. Alternative possible matches provide endogenous outside options. We ask when do such markets clear efficiently and find inefficiencies – mismatch and delay – to be pervasive. Mismatch occurs whenever an agent is at risk of losing a binding endogenous outside option. Delay occurs, instead, when the market evolves in favor of an agent. Delay can be extensive and structured with vertically differentiated markets endogenously clearing form the top down.

May 23, 2017, at 16:30 (Week 5 of Trinity):

Speaker: Francesco Nava, London School of Economics
Venue: Staircase L, Conference Room (Upstairs)
Title: Differentiated Durable Good: Monopoly and Competition

Abstract: We present results contained in two projects on differentiated durable goods. The monopoly project analyses a durable good problem in which multiple varieties can be produced and sold. A robust Coase conjecture establishes that the market must eventually clear, that profits must exceed static optimal market-clearing profits, and that profits converge to this lower bound in all stationary equilibria when prices can be revised instantaneously. Equilibrium pricing, however, is neither efficient nor competitive, as static optimal market-clearing prices are never competitive with more than one variety and seldom efficient. Similar conclusions apply when products can be scrapped albeit at possibly smaller mark-ups. The analysis also delivers insights on product design for such environments. The competition project considers scenarios in which the differentiated products are sold by competing producers. The main conclusions establish when entry by a competitor increases equilibrium profit of an incumbent in any equilibrium of the dynamic pricing game.

The talks during the 2015/6 academic year:

November 24th, 2015, at 17:00 (Week 7 of Michaelmas):

Speaker: Dieter Balkenborg, University of Exeter
Title: Rationalizability and Logical Inference

Abstract: In a model of modal propositional logic it is shown that the assumption of rationality and intelligence of the players implies that only rationalizable strategies can be played, and nothing more can be inferred. Hereby the assumption of “intelligence” refers to the familiar inference rule of necessitation and expresses that whatever an outside observer can deduce about the play of the game can be inferred by the players themselves, if they have the same information.

December 1st, 2015, at 17:00 (Week 8 of Michaelmas):

Speaker: Yuval Heller, University of Oxford
Title: Observations on Cooperation (joint with Erik Mohlin)

Abstract: We study environments in which agents are randomly matched to play a game, and before the interaction begins each agent observes a limited amount of information about the partner’s aggregate behavior. We develop a novel modelling approach for such environments and apply it to study the Prisoner’s Dilemma. We first show that defection is evolutionarily stable for any level of observability and behavioral noise. Next we classify the Prisoner’s Dilemma into four categories of games, and we fully characterize when cooperation is evolutionarily stable in each of them. Click here for paper.

January 19th, 2016, at 16:30 (Week 1 of Hillary):

Speaker: Maia King, Queen Mary University of London
Title: Whom Can You Trust? Reputation and Cooperation in Networks

Abstract. Community enforcement is an important device for sustaining efficiency in some repeated games of cooperation. Applying Dixit’s (2003b) community enforcement model in a network setting, we investigate cooperation when information about players’ reputations spreads to their future partners through links in the social network. We nd that information supports cooperation by increasing trust between players, and obtain the `radius of trust’: an endogenous network listing the potentially cooperative relationships between pairs of players in a community. We identify two aspects of trust. Players are trusted if others can communicate about them, which we link to 2-connectedness of the network and the length of cycles within it. Players are trusting if they are more likely to receive information from others through their network connections; this is linked to a new centrality measure that depends on the probabilities of information transmission between nodes. We nd a new closed-form function to identify these probabilities. Click here for paper.

February 2nd, 2016, at 16:30 (Week 3 of Hillary):

Speaker: Andrew Mell, University of Oxford
Title: Fooling Some of the People Some of the Time: Reputation Management and Optimal Betrayal

Abstract: A rational long lived player plays against a series of short lived playerswho use a variant of the Adaptive Play algorithm. In equilibrium, if the long lived player has an intermediate level of patience and the temptation to act opportunistically increases in the probability of being trusted, they will build an intermediate reputation. If the temptation to act opportunistically falls with the probability of being trusted and their initial reputation is relatively good, they will build and maintain a perfect reputation. Otherwise they will consume all the reputation they have. A public relations professional can manipulate the sampling of the short lived players to the benet of the long lived player. As a result a patient long lived player’s behavior will worsen while an impatient long lived player’s behavior may improve. Click here for paper.

February 16th, 2016, at 16:30 (Week 5 of Hillary)

Speaker: Bassel Tarbush, University of Oxford (joint w/ Francis Dennig)
Title: Fertility choice, intergenerational transmission, and the distribution of capital

Abstract: We propose a simple dynamic stochastic model of the intergenerational transmission of capital to study the impact of the fertility decision on its long run distribution. We assume that parents maximize the number of children they have subject to each child having the same economic possibilities as their parent. This yields a negative relationship between fertility and capital that is consistent with empirical observation. Combined with some degree of rivalry amongst siblings for parental resources, the main consequence is that inequality in the steady state distribution of capital is increasing in the degree of rivalry.

March 1st, 2016, at 16:30 (Week 7 of Hillary):

Speaker: Caleb Koch, ETH Zurich (joint w/ Heinrich Nax)
Title:  Evolution of a common-pool resource: theory and evidence from groundwater usage

Abstract: We apply game theory in a large-scale empirical study of the individual-level decisions and dynamics of water usage in the context of the High Plains Aquifer, one of the largest (common-pool) groundwater systems in the world. We study the involved interactions and empirically test the components that drive individual water (over-)usages, in particular focusing on behavioural heuristics and “anchoring” norms (Tversky and Kahneman 1974). Combining theory and data, we recommend an optimal quota policy that leverages individual-level behaviour and social norms with the aim of sustaining groundwater levels.

May 17th, 2016, at 16:30 (Week 4 of Trinity)

Speaker: James Best, University of Oxford
Title: Anti-Social Learning

Abstract: I examine the effect of social learning on social norms of cooperation. To this end I develop an `anti-social learning’ game. This is a dynamic social dilemma in which all agents know how to cooperate but a proportion are “informed” and know of privately profitable but socially costly, or uncooperative, actions. In equilibrium agents are able to infer, or learn, the  payoffs to the actions of prior agents. Agents can then learn through observation that some socially costly action is privately profitable. This implies that an informed agent behaving uncooperatively can induce others to behave uncooperatively when, in the absence of observational learning, they would have otherwise been cooperative. However, this influence also gives informed agents an incentive to cooperate — not cooperating may induce others to not cooperate. Within a finite horizon setting this game yields a unique equilibrium in which there is cooperation up to a finite distance from the end point of the game. In an infinite horizon setting the learning mechanism may rule out certain equilibria that would exist in the game without social learning, particularly it can rule out equilibria in which all agents defect. I use this model to give conditions under which social learning propagates cooperative behaviour and conditions under which social learning propagates uncooperative behaviour.

May 31st, 2016, at 16:30 (Week 6 of Trinity)

Speaker: Andy Zapechelnyuk, University of Glasgow (Joint work with Karl Schlag, University of Vienna)
Place: Staircase L, Conference Room (Upstairs)
Title: Robust Sequential Search

Abstract: The paper revisits the problem of sequential search with recall. An individual observes alternatives which arrive one by one. At every stage the individual has to make a choice: to stop and select one of the alternatives arrived thus far, or to continue the search.

In a classical formulation of the problem the individual knows the environment (the probability distribution according to which values of alternatives are drawn), or is able to form beliefs about possible environments. These assumptions are strong and often unrealistic. Imagine a shopper visiting one shop after another in search for a particular product. How could the shopper know the probability that the next shop will have the desired product? How would the shopper form beliefs about such probabilities?

In this paper we assume that the individual has no information about the environment and does not rely on prior beliefs when making decisions. We seek robust search rules that guarantee a “good” performance without relying on any specific assumptions about the distribution. We choose a worst case approach and look for a rule that performs as close as possible to the first-best payoff attainable if the true environment were known. We show that the individual can guarantee to get 1/4 of the first-best payoff at the outset, with the ratio increasing as better alternatives are drawn.

June 14th, 2016, at 16:30 (Week 8 of Trinity)

Speaker: Peiran Jiao, University of Oxford

Place: Staircase L, Large Lecture Room
Title: Experience-Based Belief Distortion

Abstract: People overweight experience relative to descriptive and observational information, in games, portfolio choice, etc. However, little is known about how beliefs are biased by experienced payoffs. This paper offers a simple model of experience-based belief distortion, where the decision maker with good (bad) experience misinterprets bad (good) signals, and overestimates future good (bad) states. Two experiments were conducted to test the model predictions. The first experiment asked subjects to predict future prices after viewing some stock price charts, whereas experienced gain/loss was exogenously assigned. The second experiment further provided information about the outcome-generating processes to allow for Bayesian updating as a benchmark. Subjects who gained reported significantly more optimistic guesses than those who lost after viewing the same sequence; in belief updating, they overweighted new evidence in favor of the signal from which they gained. The paper concludes with a discussion of the model implications.