Setting: agents making strategic decisions (new) in dynamic environments.

• Entry and exit: Collard-Wexler (2013)
• Sunk costs: Ryan (2012)
• Innovation: Goettler and Gordon (2011)
  • (or whatever changes in response to investment)
• Exploitation of natural resources: Huang and Smith (2014)
• Durable goods: Esteban and Shum (2007)

Lit review: forthcoming IO Handbook chapter Aguirregabiria, Collard-Wexler, and Ryan (2021)

Typically in IO we study agents in strategic environments. Complicated in dynamic environments.

• Curse of dimensionality
  • Single agent: need to track what the agent sees (\(k\) states)
  • Multi-agent: need to keep track what every agent sees (\(k^J\)states)
  • Difference exponential in the number of agents
• Expectations
  • Need not only to keep track of how the environment evolves
  • … but also of how other players act
• Equilibrium
  • Because of the strategic interaction, the Bellman equation is not a contraction anymore
    • Equilibrium existence?
    • Equilibrium uniqueness?

We will cover first the estimation and then the computation of dynamic games

• Weird…
• Standard estimation method: Bajari, Benkard, and Levin (2007)
• Does not require to solve the model
• Indeed, that’s the advantage of the method
• Disadvantages: still need to solve the model for counterfactuals
• So we’ll cover computation afterwards

Last: bridge between Structural IO and Artificial Intelligence

• Different objectives but similar methods
• Dynamic tools niche in IO but at the core of AI

Stylized version of Ericson and Pakes (1995) (no entry/exit)

• \(J\) firms (products) indexed by \(j \in \lbrace 1, ..., J \rbrace\)

• Time \(t\) is dicrete, horizon is infinite

• States \(s_{jt} \in \lbrace 1, ... \bar s \rbrace\): quality of product \(j\) in period \(t\)

• Actions \(a_{jt} \in \mathbb R^+\): investment decision of firm \(j\) in period \(t\)

• Static payoffs \[ \pi_j (s_{jt}, \boldsymbol s_{-jt}, a_{jt}; \theta^\pi) \] where

  • \(\boldsymbol s_{-it}\): state vector of all other firms in period \(t\)
  • \(\theta^\pi\): parameters that govern static profits

Note: if we micro-fund \(\pi(\cdot)\) , e.g. with some demand and supply model, we have 2 strategic decisions: prices (static) and investment (dynamic).

• State transitions \[ \boldsymbol s_{t+1} = f(\boldsymbol s_t, \boldsymbol a_t, \boldsymbol \epsilon_t; \theta^f) \] where

  • \(\boldsymbol a_t\): vector of actions of all firm
  • \(\boldsymbol \epsilon_t\): vector of idiosyncratic shocks
  • \(\theta^f\): parameters that govern state transitions

• Objective function: firms maximize expected discounted future profits \[ \max_{\boldsymbol a} \ \mathbb E_t \left[ \sum_{\tau=0}^\infty \beta^{\tau} \pi_{j, t+\tau} (\theta^\pi) \right] \]

The value function of firm \(j\) at time \(t\) in state \(\boldsymbol s_{t}\), under a set of strategy functions \(\boldsymbol P\) (one for each firm) is \[ V^{\boldsymbol P_{-j}}_{j} (\mathbf{s}_{t}) = \max_{a_{jt} \in \mathcal{A}_j \left(\mathbf{s}_{t}\right)} \Bigg\lbrace \pi_{j}^{\boldsymbol P_{-j}} (a_{jt}, \mathbf{s}_{t} ; \theta^\pi ) + \beta \mathbb E_{\boldsymbol s_{t+1}} \Big[ V_{j}^{\boldsymbol P_{-j}} \left(\mathbf{s}_{t+1}\right) \ \Big| \ a_{jt}, \boldsymbol s_{t} ; \theta^f \Big] \Bigg\rbrace \] where

• \(\pi_{j}^{\boldsymbol P_{-j}} (a_{jt}, \mathbf{s}_{t} ; \theta^\pi )\) are the static profits of firm \(j\) given action \(a_{jt}\) and policy functions \(\boldsymbol P_{-j}\) for all firms a part from \(j\)

• The expecation \(\mathbb E\) is taken with respect to the conditional transition probabilities \(f^{\boldsymbol P_{-j}} (\mathbf{s}_{t+1} | \mathbf{s}_{t}, a_{jt} ; \theta^f)\)

Equillibrium notion: Markow Perfect Equilibrium (Maskin and Tirole 1988)

• Assumption: players’ strategies at period \(t\) are functions only of payoff-relevant state variables at the same period
• Definition: a set of \(J\) value and policy functions, \(\boldsymbol V\) and \(\boldsymbol P\) such that each firm
  1. maximizes its value function \(V_j\)
  2. given the policy function of every other firm \(\boldsymbol P_{-j}\)

What is it basically?

• Nash Equilibrium in the policy functions
• What are we ruling out?
  • Strategies that depend on longer histories
  • E.g. “has anyone ever cheated in a cartel?”

We want to estimate 2 sets of parameters:

• \(\theta^\pi\): parameterizes period profit function \(\pi(\cdot)\)
• \(\theta^f\): parameterizes state transition function \(f(\cdot)\)

Generally 2 approaches

1. Full solution
   • Impractical (we’ll see more details later)
2. Rely on some sort of Hotz and Miller (1993) CCP inversion
   • Aguirregabiria and Mira (2007)
   • Bajari, Benkard, and Levin (2007)
   • Pakes, Ostrovsky, and Berry (2007)
   • Pesendorfer and Schmidt-Dengler (2008)

Bajari, Benkard, and Levin (2007) plan

1. Estimate transition probabilities and conditional choice probabilities from the data
2. Use them to simulate the expected value function, given a set of parameters
3. Use optimality of estimated choices to pin down static profit parameters
   • I.e. repeat (2) for alternative strategies
     • By definition suboptimal
   • Estimating equation: values implied by observed strategies should be higher than values implied by alternative strategies

• Estimate the transition probabilities \(f ( \cdot | a_{jt}, \boldsymbol s_t; \hat \theta^f )\)
  • I.e. what is the observed frequency of any state-to-state transition?
  • For any given action of firm \(j\)
• … and conditional choice probabilities \(\hat P_j(\cdot | \boldsymbol s_t)\)
  • I.e. what is the probability of each action, for each firm \(j\) in each state \(\boldsymbol s\)
• Can be done non-parametrically
  • i.e. just observe frequencies
  • Conditional on having enough data
  • Note: need to estimate transitions, conditional on each state and action
  • Problem with many states and actions, but especially with many players
    • Curse of dimensionality
    • Number of states increases exponentially in number of players

Important: parametric assumptions would contradict the model for the estimation of value/policy functions

First step: from transitions \(f(\hat \theta^f)\) and CCPs \(\boldsymbol{\hat P}\) to values

• We can use transitions and CCPs to simulate histories (of length \(\tilde T\))

  • of states \(\lbrace \boldsymbol{\tilde{s}_{\tau}} \rbrace_{\tau = 1}^{\tilde T}\)
  • and actions \(\lbrace \boldsymbol{\tilde{a}_{\tau}} \rbrace_{\tau = 1}^{\tilde T}\)

• Given a parameter value \(\tilde \theta^\pi\), we can compute static payoffs: \(\pi_{j}^{\boldsymbol {\hat{P}_{-j}}} \left( \tilde a_{j\tau}, \boldsymbol{\tilde s}_{\tau} ; \tilde \theta^\pi \right)\)

• Simulated history + static payoffs = simulated value function \[ {V}_{j}^{\boldsymbol {\hat{P}}} \left(\boldsymbol{s}_{t} ; \tilde \theta^\pi \right) = \sum_{\tau=0}^{\tilde T} \beta^{\tau} \pi_{j}^{\boldsymbol {\hat{P}_{-j}}} \left( \tilde a_{j\tau}, \boldsymbol{\tilde s}_{\tau} ; \tilde \theta^\pi \right) \]

• We can average over many, e.g. \(R\), simulated value functions to get an expected value function \[ {V}_{j}^{\boldsymbol {\hat{P}}, R} \left( \boldsymbol{s}_{t} ; \tilde \theta^\pi \right) = \frac{1}{R} \sum_{r=0}^{R}\Bigg( \sum_{\tau=0}^{\tilde T} \beta^{\tau} \pi_{j}^{\boldsymbol {\hat{P}_{-j}}} \left(\tilde a^{(r)}_{j\tau}, \boldsymbol{\tilde s}^{(r)}_{\tau} ; \tilde \theta^\pi \right) \Bigg) \]

For \(r = 1, ..., R\) simulations do:

• Initialize firms value to zero
• Fot \(\tau=0, ..., \tilde T\) do
  • For each state in \(\boldsymbol{\tilde s}^{(r)}_{\tau}\) do:
    • Use \(\boldsymbol{\hat P}\) to draw a vector of firm actions \(\boldsymbol{\tilde a}^{(r)}_{\tau}\)
    • For each firm \(j = 1, ..., J\) do:
      • Compute static profits \(\pi_{j}^{\boldsymbol {\hat{P}_{-j}}} \left(\tilde a^{(r)}_{j\tau}, \boldsymbol{\tilde s}^{(r)}_{\tau} ; \tilde \theta^\pi \right)\)
      • Add discounted profits \(\beta^{\tau} \pi_{j}^{\boldsymbol {\hat{P}_{-j}}} \left(\tilde a^{(r)}_{j\tau}, \boldsymbol{\tilde s}^{(r)}_{\tau} ; \tilde \theta^\pi \right)\) to the value function
    • Use \(f ( \cdot | \boldsymbol {a_{t}}, \boldsymbol s_t; \hat \theta^f )\) to draw the next state \(\boldsymbol{\tilde s}^{(r)}_{\tau + 1}\)
    • Use the next state, \(\boldsymbol{\tilde s}^{(r)}_{\tau + 1}\) as current state for the next iteration

Then average all the value functions together to obtain an expected value function \(V_{j}^{\boldsymbol {\hat{P}}, R} \left(\boldsymbol{s}_{t} ; \tilde \theta^\pi \right)\)

Note: advantage of simulations: can be parallelized

What have we done so far?

• Given some parameters \(\theta^\pi\), we computed the expected value function

How do we pick the \(\theta^\pi\) that best rationalizes the data?

• I.e. what is the objective function?
• Potentially many options

BBL idea

• the expected value function has to be optimal, given the CCPs
• I.e. any other policy function should give a lower expected value
• “Best” \(\theta^\pi\): those for which the implied expected value function under the estimated CCPs is greater than the one implied by any other CCP

Note: it’s an inequality statement

Idea

• If the observed policy \({\color{green}{\boldsymbol{\hat P}}}\) is optimal,

  • All other policies \({\color{red}{\boldsymbol{\tilde P}}}\)

  • … at the true parameters \(\theta^f\)

  • … should give a lower expected value \[ V_{j}^{{\color{red}{\boldsymbol{\tilde P}}}, R} \left( \boldsymbol{s}_{t} ; \tilde \theta^\pi \right) \leq V_{j}^{{\color{green}{\boldsymbol{\hat P}}}, R} \left( \boldsymbol{s}_{t} ; \tilde \theta^\pi \right) \]

• So which are the true parameters?

  • Those for which any deviation from the observed policy \({\color{green}{\boldsymbol{\hat P}}}\) yields a lower value

  • Objective function to minimize: violations under alternative policies \({\color{red}{\boldsymbol{\tilde P}}}\) \[ \min_{\tilde \theta^\pi} \sum_{\boldsymbol s_{t}} \sum_{{\color{red}{\boldsymbol{\tilde P}}}} \Bigg[\min \bigg\lbrace V_{j}^{{\color{green}{\boldsymbol{\hat P}}}, R} \left( \boldsymbol{s}_{t} ; \tilde \theta^\pi \right) - V_{j}^{{\color{red}{\boldsymbol{\tilde P}}}, R} \left( \boldsymbol{s}_{t} ; \tilde \theta^\pi \right) \ , \ 0 \bigg\rbrace \Bigg]^{2} \]

Estimator: \(\theta^\pi\) that minimizes the average (squared) magnitude of violations for any alternative policy \({\color{red}{\boldsymbol{\tilde P}}}\) \[ \hat{\theta}^\pi= \arg \min_{\tilde \theta^\pi} \sum_{\boldsymbol s_{t}} \sum_{{\color{red}{\boldsymbol{\tilde P}}}} \Bigg[\min \bigg\lbrace V_{j}^{{\color{green}{\boldsymbol{\hat P}}}, R} \left( \boldsymbol{s}_{t} ; \tilde \theta^\pi \right) - V_{j}^{{\color{red}{\boldsymbol{\tilde P}}}, R} \left( \boldsymbol{s}_{t} ; \tilde \theta^\pi \right) \ , \ 0 \bigg\rbrace \Bigg]^{2} \]

• \(\min \Big\lbrace V_{j}^{{\color{green}{\boldsymbol{\hat P}}}, R} - V_j^{{\color{red}{\boldsymbol{\tilde P}}}, R} \ , \ 0 \Big\rbrace\) to pick only the violations
  • If \({\color{green}{\boldsymbol{\hat P}}}\) implies higher value, we can ignore
  • Doesn’t matter by how much you respect the inequality
• Which alternative policies \({\color{red}{\boldsymbol{\tilde P}}}\) should we use?
  • In principle, any perturbation is ok
  • But in practice, if we perturbe it too much, we can go too far off
  • Tip 1: start with very small perturbations
  • Tip 2: use perturbation that sensibly affect the dynamics
    • E.g. exiting in a state in which a firm is not a competitive threat
  • Tip 3: use perturbations on dimensions that are relevant for the research question
    • E.g. they affect dimensions where you want to make counterfactual predictions

We have seen that there are competing methods.

What are the advantages of Bajari, Benkard, and Levin (2007) over those?

1. Continuous actions
   • BBL does not require actions to be discretised
   • You can just sample actions from the data!
2. Choice of alternative CCPs
   • The researcher is free to choose the alternative CCPs \({\color{red}{\boldsymbol{\tilde P}}}\)
   • Pros: can make source of variation more transparent
     • allows the researcher to focus on those predictions of the model that are key for the specific research questions
   • Cons: it’s a very high dimensional space
     • There are very very many alternative policy functions

1. Computational curse of dimensionality is gone (in the state space)
   • But we have a curse of dimensionality in data
   • Need a lot of markets because now 1 market is 1 observation
2. Multiple equilibria??
   • We are basically assuming it away
   • Estimating the CCPs in the first stage we assume that is the equilibrium that is played in all markets at all times
   • To run counterfactuals, we still need to solve the model
3. Unobserved heterogeneity
   • Kasahara and Shimotsu (2009): how to identify the (minimum) number of unobserved types
   • Arcidiacono and Miller (2011): how to use an EM algorithm for the 1st stage estimation with unobserved types, conditional on the number of types
   • Berry and Compiani (2021): instrumental variables approach, relying on observed states in the distant past
4. Non-stationarity
   • If we have a long time period, something fundamentally might have changed

Ericson and Pakes (1995) and companion paper Pakes and McGuire (1994) for the computation

• \(J\) firms indexed by \(j \in \lbrace 1, ..., J \rbrace\)

• Time \(t\) is dicrete \(t\), horizon is infinite

• State \(s_{jt}\): quality of firm \(j\) in period \(t\)

• Per period profits \[ \pi (s_{jt}, \boldsymbol s_{-jt}, ; \theta^\pi) \] where

  • \(\boldsymbol s_{-it}\): state vector of all other firms in period \(t\)
  • \(\theta^\pi\): parameters that govern static profits

• We can micro-fund profits with some demand and supply functions

  • There can be some underlying static strategic interaction
  • E.g. logit demand and bertrand competition in prices \(p_{it}\)

Investment: firms can invest an dollar amount \(x\) to increase their future quality

• Continuous decision variable (\(\neq\) Rust)

• Probability that investment is successful \[ \Pr \big(i_{jt} \ \big| \ a_{it} = x \big) = \frac{\alpha x}{1 + \alpha x} \]

• Higher investment, higher success probability

• \(\alpha\) parametrizes the returns on investment

Quality depreciation

• With probability \(\delta\), quality decreases by one level

Law of motion \[ s_{j,t+1} = s_{jt} + i_{jt} - \delta \]

Note that in Ericson and Pakes (1995) we have two separate decision variables

1. Static decision variable: price \(p_{jt}\)
2. Dynamic decision variable: investment \(i_{jt}\)

Does not have to be the case!

Example: Besanko et al. (2010)

• Model of learning-by-doing: firms decrease their marginal cost through sales
• State variable: firm stock of know how \(e\)
  • The higher the stock of know-how, the lower the marginal cost
  • Increases when a firm manages to make a sale
    • \(q \in [0,1]\) now is both static quantity and transition probability
• Single decision variable: price \(p\)
  • Usual static effects on profits \(\pi_{jt} = (p_{jt} - c(e_{jt})) \cdot q_j(\boldsymbol p_t)\)
  • But also dynamic effect through transition probabilities
    • Probability of increasing \(e_t\): \(q_j(\boldsymbol p_t)\)

Firms maximize the expected flow of discounted profits \[ \max_{\boldsymbol a} \ \mathbb E_t \left[ \sum_{\tau=0}^\infty \beta^{\tau} \pi_{j, t+\tau} (\theta^\pi) \right] \] Markow Perfect Equilibrium

Equillibrium notion: Markow Perfect Equilibrium (Maskin and Tirole 1988)

• A set of \(J\) value and policy functions, \(\boldsymbol V\) and \(\boldsymbol P\) such that each firm
  1. maximizes its value function \(V_j\)
  2. given the policy function of every other firm \(\boldsymbol P_{-j}\)

One important extension is exit.

• In each time period, incuments decide whether to stay
• … or exit and get a scrap value \(\phi^{exit}\)

The Belman Equation of incumbent \(j\) at time \(t\) is \[ V^{\boldsymbol P_{-j}}_{j} (\mathbf{s}_{t}) = \max_{d^{exit}_{jt} \in \lbrace 0, 1 \rbrace} \Bigg\lbrace \begin{array}{c} \beta \phi^{exit} \ , \newline \max_{a_{jt} \in \mathcal{A}_j \left(\mathbf{s}_{t}\right)} \Big\lbrace \pi_{j}^{\boldsymbol P_{-j}} (a_{jt}, \mathbf{s}_{t} ; \theta^\pi ) + \beta \mathbb E_{\boldsymbol s_{t+1}} \Big[ V_{j}^{\boldsymbol P_{-j}} \left(\mathbf{s}_{t+1}\right) \ \Big| \ a_{jt}, \boldsymbol s_{t} ; \theta^f \Big] \Big\rbrace \end{array} \Bigg\rbrace \] where

• \(\phi^{exit}\): exit scrap value
• \(d^{exit}_{jt} \in \lbrace 0,1 \rbrace\): exit decision

We can also incorporate endogenous entry.

• One or more potential entrants exist outside the market
• They can pay an entry cost \(\phi^{entry}\) and enter the market at a quality state \(\bar s\)
• … or remain outside at no cost

Value function \[ V_{j}^{\boldsymbol P_{-j}} (e, \boldsymbol x_{-jt} ; \theta) = \max_{d^{entry} \in \lbrace 0,1 \rbrace } \Bigg\lbrace \begin{array}{c} 0 \ ; \newline - \phi^{entry} + \beta \mathbb E_{\boldsymbol s_{t+1}} \Big[ V_{j}^{\boldsymbol P_{-j}} (\bar s, \boldsymbol s_{-j, t+1} ; \theta) \ \Big| \ \boldsymbol s_{t} ; \theta^f \Big] \end{array} \Bigg\rbrace \] where

• \(d^{entry} \in \lbrace 0,1 \rbrace\): entry decision
• \(\phi^{entry}\): entry cost
• \(\bar s\): state in which entrants enters (could be random)

Do we observe potential entrants?

• Igami (2017): tech industry announce their entry
• Critique: not really potential entrants, they are half-way inside

Doraszelski and Satterthwaite (2010): a MPE might not exist in Ericson and Pakes (1995) model.

Solution

• Replace fixed entry costs \(\phi^{entry}\) and exit scrap values \(\phi^{exit}\) with random ones
• It becomes a game of incomplete information
  • First explored in Rust (1994)
• New equilibrium concept

Markov Perfect Bayesian Nash Equilibrium (MPBNE)

• Basically the same, with rational beliefs on random variables

Solving the model is very similar to Rust

• Given parameter values \(\theta\)
• Start with a guess for the value and policy functions
• Until convergence, do:
  • For each firm \(j = 1, ..., J\), do:
    • Take the policy functions of all other firms
    • Compute the implied transition probabilities
    • Use them to compute the new policy function for firm \(j\)
    • Compute the implied value function

Where do things get complicated / tricky? Policy function update

Imagine a stylized exit game with 2 firms

• Easy to get an update rule of the form: “exit if opponent stays, stay if opponent exits”

Computationally

• Initialize policy functions to \((exit, exit)\)
• Iteration 1:
  • Each firm takes opponent policy as given: \(exit\)
  • Update own optimal policy: \(stay\)
  • New policy: \((stay, stay)\)
• Iteration 2: \((stay, stay) \to (exit, exit)\)
• Iteration 2: \((exit, exit) \to (stay, stay)\)
• Etc…

Issues: value function iteration might not converge and equilibrium multeplicity.

• Try different starting values
  • Often it’s what makes the biggest difference
  • Ideally, start as close as possible to true values
  • Approximation methods can help (we’ll see more later)
    • I.e. get a fast approximation to use as starting vlaue for solution algorithm
• Partial/stochastic value function update rule
  • Instead of \(V' = T(V)\), use \(V' = \alpha T(V) + (1-\alpha)V\)
  • Very good to break loops, especially if \(\alpha\) is stochastic, e.g. \(\alpha \sim U(0,1)\)
• How large is the support of the entry/exit costs?
  • If support is too small, you end up back in the entry/exit loop
• Try alternative non-parallel updating schemes
  • E.g. update value one state at the time (in random order?)
• Last but not least: change the model
  • In particular, from simultaneous to alternating moves
  • or continuous time

How to find them?

• Besanko et al. (2010) and Borkovsky, Doraszelski, and Kryukov (2010): homotopy method
  • can find some equilibria, but not all
  • complicated to implement: need to compute first order conditions \(H(\boldsymbol V, \theta) = 0\) and their Jacobian \(\Delta H(\boldsymbol V, \theta)\)
  • Idea: trace the equilibrium correspondence \(H^{-1} = \lbrace (\boldsymbol V, \theta) : H(\boldsymbol V, \theta) = 0 \rbrace\) in the value-parameter space
• Eibelshäuser and Poensgen (2019)
  • Markov Quantal Response Equilibrium
  • approact dynamic games from a evolutionary game theory perspective
    • actions played at random and those bringing highest payoffs survive
  • \(\to\) homothopy method guaranteed to find one equilibrium
• Pesendorfer and Schmidt-Dengler (2010): some equilibria are not Lyapunov-stable
  • BR iteration cannot find them unless you start exactly at the solution
• Su and Judd (2012) and Egesdal, Lai, and Su (2015): same point, but numerically
  • using MPEC approach

Can we assume them away?

• Igami (2017)
  • Finite horizon
  • Homogenous firms (in profit functions and state transitions)
  • One dynamic move per period (overall, not per-firm)
• Abbring and Campbell (2010)
  • Entry/exit game
  • Homogeneous firms
  • Entry and exit decisions are follow a last-in first-out (LIFO) structure
    • “An entrant expects to produce no longer than any incumbent”
• Iskhakov, Rust, and Schjerning (2016)
  • can find all equilibria, but for very specific class of dynamic games
  • must always proceed “forward”
    • e.g. either entry or exit but not both
  • Idea: can solve by backward induction even if horizon is infinite

What are the computational bottlenecks? \[ V^{\boldsymbol P_{-j}}_{j} ({\color{red}{\mathbf{s}_{t}}}) = \max_{a_{jt} \in \mathcal{A}_j \left(\mathbf{s}_{t}\right)} \Bigg\lbrace \pi_{j}^{\boldsymbol P_{-j}} (a_{jt}, \mathbf{s}_{t} ; \theta^\pi ) + \beta \mathbb E_{{\color{red}{\mathbf{s}_{t+1}}}} \Big[ V_{j}^{\boldsymbol P_{-j}} \left(\mathbf{s}_{t+1}\right) \ \Big| \ a_{jt}, \boldsymbol s_{t} ; \theta^f \Big] \Bigg\rbrace \]

1. Dimension of the state space
   • In single agent problems, we have as many states as many values of \(s_{jt}\) (\(k\))
   • In dynamics games, the state space goes from \(k\) to \(k^J\)
   • symmetry helps: state \([1,2,3]\) and \([1,3,2]\) become the same for firm 1
   • How much do we gain? From \(k^J\) to \(k \cdot {k + J - 2 \choose k - 1}\)
2. Dimension of the integrand
   • If in single agent problems, we have to integrate over \(\kappa\) outcomes,
     • 4 in Rust: engine replaced (yes|no) \(\times\) mileage increases (yes|no)
   • … in dynamic games, we have to consider \(\kappa^J\) outcomes

Note: bottlenecks are not addittive but multiplicative: have to solve the expectation for each point in the state space. Improving on any of the two helps a lot.

Two and a half classes of solutions:

1. Computational: approximate the equilibrium
   • Doraszelski (2003): use Chebyshev polynomials for a basis function
   • Farias, Saure, and Weintraub (2012): combine approximations with a MPEC-like approach
2. Conceptual: define another game
   • Weintraub, Benkard, and Van Roy (2008): oblivious equilibrium
   • Ifrach and Weintraub (2017): moment based equilibrium
   • Doraszelski and Judd (2012): games in continuous time
   • Doraszelski and Judd (2019): games with random moves
3. Kind of both: Pakes and McGuire (2001)
   • experience-based equilibrium (Fershtman and Pakes 2012)

Note: useful also to get good starting values for a full solution method!

Weintraub, Benkard, and Van Roy (2008): what if firms had no idea about the state of other firms?

• or atomistic firms

The value function becomes \[ V_{j} ({\color{red}{s_{t}}}) = \max_{a_{jt} \in \mathcal{A}_j \left({\color{red}{s_{t}}}\right)} \Bigg\lbrace {\color{red}{\mathbb E_{\boldsymbol s_t}}} \Big[ \pi_{j} (a_{jt}, \mathbf{s}_{t} ; \theta^\pi ) \Big| P \Big] + \beta \mathbb E_{{\color{red}{s_{t+1}}}} \Big[ V_{j} \left({\color{red}{s_{t+1}}}\right) \ \Big| \ a_{jt}, {\color{red}{s_{t}}} ; \theta^f \Big] \Bigg\rbrace \]

• Now the state is just \(s_t\) instead of \(\boldsymbol s_t\)
  • Huge computational gain: from \(k^J\) points to \(k\)
  • Also the expectation of future states is taken over \(3\) instead of \(3^J\) points
    • (3 because quality can go up, down or stay the same)
• But need to compute static profits as the expected value given the current policy function
  • Need to keep track of the asymptotic state distribution as you iterate the value

Doraszelski and Judd (2019): what if instead of simultaneously, firms would move one at the time at random?

• Important: to have the same frequency of play, a period now is \(J\) times shorter

The value function becomes \[ V^{\boldsymbol P_{-j}}_{j} (\mathbf{s}_{t}, {\color{red}{n=j}}) = \max_{a_{jt} \in \mathcal{A}_j \left(\mathbf{s}_{t}\right)} \Bigg\lbrace {\color{red}{\frac{1}{J}}}\pi_{j}^{\boldsymbol P_{-j}} (a_{jt}, \mathbf{s}_{t} ; \theta^\pi ) + {\color{red}{\sqrt[J]{\beta}}} \mathbb E_{{\color{red}{n, s_{j, t+1}}}} \Big[ V_{j}^{\boldsymbol P_{-j}} \left(\mathbf{s}_{t+1}, {\color{red}{n}} \right) \ \Big| \ a_{jt}, \boldsymbol s_{t} ; \theta^f \Big] \Bigg\rbrace \]

• \(n\): indicates whose turn is to play
• since a turn is \(J\) times shorter, profits are \(\frac{1}{J} \pi\) and discount factor is \(\sqrt[J]{\beta}\)

Computational gain

• Expectation now taken over \(n, s_{j, t+1}\) instead of \(\boldsymbol s_{t+1}\)
• I.e. \(Jk\) points instead of \(3^k\) (3 because quality can go up, down or stay the same)
• Huge computational difference!

Doraszelski and Judd (2012): what’s the advantage of continuous time?

• Probability that two firms take a decision simultaneously is zero

With continuous time, the value function becomes \[ V^{\boldsymbol P_{-j}}_{j} (\mathbf{s}_{t}) = \max_{a_{jt} \in \mathcal{A}_j \left(\mathbf{s}_{t}\right)} \Bigg\lbrace \frac{1}{\lambda(a_{jt}) - \log(\beta)} \Bigg( \pi_{j}^{\boldsymbol P_{-j}} (a_{jt}, \mathbf{s}_{t} ; \theta^\pi ) + \lambda(a_{jt}) \mathbb E_{\boldsymbol s_{t+1}} \Big[ V_{j}^{\boldsymbol P_{-j}} \left(\mathbf{s}_{t+1}\right) \ \Big| \ a_{jt}, \boldsymbol s_{t} ; \theta^f \Big] \Bigg) \Bigg\rbrace \]

• \(\lambda(a_{jt}) = \delta + \frac{\alpha a_{jt}}{1 + \alpha a_{jt}}\) is the hazard rate for firm \(j\) that something happens
  • i.e. either an increase in quality, with probability \(\frac{\alpha a_{jt}}{1 + \alpha a_{jt}}\)
  • … or a decrease in quality with probability \(\delta\)

Computational gain

• Now the expectation over future states \(\mathbb E_{\boldsymbol s_{t+1}}\) is over \(2J\) points instead of \(3^J\)
  • 3 because quality can go up, down or stay the same
  • 2 because in continuous time we don’t care if the state does not change (investment fails)

Which method is best?

I compare them in Courthoud (2020)

• Fastest: Weintraub, Benkard, and Van Roy (2008)
  • Effectively transforms the game into single-agent dynamics
• Best trade-off: Doraszelski and Judd (2019)
  • Simple, practical and also helps in terms of equilibrium multeplicity
• Also in Courthoud (2020): games with random order
  • Better approximation than Doraszelski and Judd (2019)
  • And similar similar time

Some applications of these methods include

• Approximation methods
  • Sweeting (2013): product repositioning among radio stations
  • Barwick and Pathak (2015): entry and exit in the real estate brokerage industry
• Oblivious equilibrium
  • Xu and Chen (2020): R&D investment in the Korean electric motor industry
• Moment based equilibrium
  • Jeon (2020): demand learning in the container shipping industry
  • Caoui (2019): technology adoption with network effects in the movie industry
  • Vreugdenhil (2020): search and matching in the oil drilling industry
• Games in continuous time
  • Arcidiacono et al. (2016): entry, exit and scale decisions in retail competition
• Games with random moves
  • Igami (2017): innovation, entry, exit in the hard drive industry

There is one method to approximate the equilibrium in dynamic games that is a bit different from the others: Pakes and McGuire (2001)

• Idea: approximate the value function by Monte-Carlo simulation
• Firms start with a guess for the alternative-specific value function
• Act according to it
• Observe realized payoffs and state transitions
• And update the alternative-specific value function according to the realized outcomes

Experience-Based Equilibrium

• Defined in Fershtman and Pakes (2012)
• Def: policy is optimal given beliefs of state transitions and observed transitions are consistent with the beliefs
• Note: definition silent on off-equilibrium path beliefs

• Players start with alternative-specific value function

  • yes, the ASV from Rust (1994)
  • \(\bar V_{j,a}^{(0)} (\boldsymbol s ; \theta)\): initial value of player \(j\) for action \(a\) in state \(\boldsymbol s\)

• Until convergence, do:

  • Compute optimal action, given \(\bar V_{j, a}^{(t)} (\boldsymbol s ; \theta)\) \[ a^* = \arg \max_a \bar V_{j, a}^{(t)} (\boldsymbol s ; \theta) \]

  • Observe the realized payoff \(\pi_{j, a^*}(\boldsymbol s ; \theta)\) and the realized next state \(\boldsymbol {s'}(\boldsymbol s, a^*; \theta)\)

  • Update the alternative-specific value function of the chosen action \(k^*\) \[ \bar V_{j, a^*}^{(t+1)} (\boldsymbol s ; \theta) = (1-\alpha_{\boldsymbol s, t}) \bar V_{j, a^*}^{(t)} (\boldsymbol s ; \theta) + \alpha_{\boldsymbol s, t} \Big[\pi_{j, a^*}(\boldsymbol s ; \theta) + \arg \max_a \bar V_{j, a}^{(t)} (\boldsymbol s' ; \theta) \Big] \] where

    • \(\alpha_{\boldsymbol s, t} = \frac{1}{\text{number of times state } \boldsymbol s \text{ has been visited}}\)

Where is the strategic interaction?

• Firm always take “best action so far” in each state
  • Start to take a new action only when the previous best has performed badly for many periods
• Remindful of literature of evolutionary game theory

Importance of starting values

• Imagine, all payoffs are positive but value initialized to zero
• First action in each state \(\to\) only action ever taken in that state
• Loophole.
  • Why? Firms always take \(\arg \max_a \bar V_a\) and never explore the alternatives

Convergence by desing

• As \(\lim_{t \to \infty} \alpha_{\boldsymbol s, t} = 1\)
• Firms stop updating the value by design

Computer Science reinforcement learning literature (AI): Q-learning

Differences

• \(\bar V_a( \boldsymbol s)\) called \(Q_a(\boldsymbol s)\), hence the name
• Firms don’t always take the optimal action
  • At the beginning of the algorithm: exploration
    • Firms take actions at random
    • Just to explore what happens taking different actions
  • Gradually shift towards exploitation
    • I.e. take the optimal action, given \(\bar V^{(t)}( \boldsymbol s)\) at iteration \(t\)
    • I.e. shift towards Pakes and McGuire (2001)

• Doraszelski, Lewis, and Pakes (2018)
  • Firm do actually learn by trial and error
  • Setting: demand learning in the UK frequency response market (electricity)
• Asker et al. (2020)
  • Uses Pakes and McGuire (2001) for estimation
  • Setting: dynamic timber auctions with information sharing
• Calvano et al. (2020)
  • Study Q-learning pricing algorithms
  • In repeated price competition with differentiated products
  • (Computational) lab experiment: what do these algorithms converge to?
  • Finding: algorithms learn reward-punishment collusive strategies