DOCSLIB.ORG

Irreversibility, Uncertainty, and Investment." Pindyck, Robert S

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Irreversibility, Uncertainty, and Investment. Published as: “Irreversibility, Uncertainty, and Investment." Pindyck, Robert S. Journal of Economic Literature Vol. 29, No. 3 (1991): 1110-1148. Copyright & Permissions Copyright © 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012, 2013, 2014, 2015, 2016, 2017, 2018 by the American Economic Association. Permission to make digital or hard copies of part or all of American Economic Association publications for personal or classroom use is granted without fee provided that copies are not distributed for profit or direct commercial advantage and that copies show this notice on the first page or initial screen of a display along with the full citation, including the name of the author. Copyrights for components of this work owned by others than AEA must be honored. Abstracting with credit is permitted. The author has the right to republish, post on servers, redistribute to lists and use any component of this work in other works. For others to do so requires prior specific permission and/or a fee. Permissions may be requested from the American Economic Association Administrative Office by going to the Contact Us form and choosing "Copyright/ Permissions Request" from the menu. Copyright © 2018 AEA Journal of Economic Literature Vol. XXIX (September 1991), pp. 1110-1148 Irreversibility,Uncertainty, and Investment By ROBERT S. PINDYCK MassachusettsInstitute of Technology My thanks to Prabhat Mehtafor his research assistance, and to Ben Bernanke, Vittorio Corbo, Nalin Kulatilaka,Robert McDonald, Louis Serven, Andreas Solimano, and two anonymous referees for helpful commentsand suggestions. Financial support was provided by MIT's Center for Energy Policy Research, by the World Bank, and by the National Science Foundation under Grant No. SES-8618502. Any errors are mine alone. I. Introduction ment expenditure can profoundly affect the decision to invest. It also undermines DESPITE ITS IMPORTANCE to economic the theoretical foundation of standard growth and market structure, the neoclassical investment models, and in- investment behavior of firms, industries, validates the net present value rule as and countries remains poorly under- it is usually taught to students in business stood. Econometric models have had school: "Invest in a project when the limited success in explaining and predict- present value of its expected cash flows ing changes in investment spending, and is at least as large as its cost." This rule we lack a clear explanation of why some and models based on it-are incorrect countries or industries invest more than when investments are irreversible and others. decisions to invest can be postponed. One problem with existing models is Irreversibility may have important im- that they ignore two important character- plications for our understanding of aggre- istics of most investment expenditures. gate investment behavior. It makes in- First, the expenditures are largely irre- vestment especially sensitive to various versible; that is, they are mostly sunk forms of risk, such as uncertainty over costs that cannot be recovered. Second, the future product prices and operating the investments can be delayed, giving costs that determine cash flows, uncer- the firm an opportunity to wait for new tainty over future interest rates, and un- information to arrive about prices, costs, certainty over the cost and timing of the and other market conditions before it investment itself. Irreversibility may commits resources. therefore have implications for macro- As an emerging literature has shown, economic policy; if a goal is to stimu- the ability to delay an irreversible invest- late investment, stability and credibility 1110 This content downloaded from 18.7.29.240 on Wed, 5 Nov 2014 09:36:58 AM All use subject to JSTOR Terms and Conditions Pindyck: Irreversibility, Uncertainty, and Investment 1111 could be much more important than tax tential competitors. (Richard Gilbert incentives or interest rates. 1989 surveys the literature on strategic What makes an investment expendi- aspects of investment.) But in most cases, ture a sunk cost and thus irreversible? delay is at least feasible. There may be Usually it is the fact that the capital is a cost to delay-the risk of entry by other firm or industry specific, that is, it cannot firms, or simply foregone cash flows-but be used productively by a different firm this cost must be weighed against the or in a different industry. For example, benefits of waiting for new information. most investments in marketing and ad- An irreversible investment opportu- vertising are firm specific, and hence are nity is much like a financial call option. clearly sunk costs. A steel plant is indus- A call option gives the holder the right, try specific-it can only be used to pro- for some specified amount of time, to pay duce steel. Although in principle the an exercise price and in return receive plant could be sold to another steel com- an asset (e.g., a share of stock) that has pany, its cost should be viewed as mostly some value. Exercising the option is ir- sunk, particularly if the industry is com- reversible; although the asset can be sold petitive. The reason is that the value of to another investor, one cannot retrieve the plant will be about the same for all the option or the money that was paid firms in the industry, so there is likely to exercise it. A firm with an investment to be little gained from selling it. (If the opportunity likewise has the option to price of steel falls so that a plant turns spend money (the "exercise price") now out, ex post, to have been a "bad"invest- or in the future, in return for an asset ment, it will also be viewed as a bad in- (e.g., a project) of some value. Again, vestment by other steel companies, so the asset can be sold to another firm, that the ability to sell the plant will not but the investment is irreversible. As be worth much.) with the financial call option, this option Even investments that are not firm or to invest is valuable in part because the industry specific are often partly irre- future value of the asset obtained by in- versible because of the "lemons" prob- vesting is uncertain. If the asset rises in lem. For example, office equipment, value, the net payoff from investing rises. cars, trucks, and computers are not in- If it falls in value, the firm need not in- dustry specific, but have resale value well vest, and will only lose what it spent to below their purchase cost, even if new. obtain the investment opportunity. Irreversibility can also arise because of How do firms obtain investment op- government regulations or institutional portunities? Sometimes they result from arrangements. For example, capital con- patents, or ownership of land or natural trols may make it impossible for foreign resources. More generally, they arise (or domestic) investors to sell assets and from a firm's managerial resources, tech- reallocate their funds. And investments nological knowledge, reputation, market in new workers may be partly irreversible position, and possible scale, all of which because of high costs of hiring, train- may have been built up over time, and ing, and firing. which enable the firm to productively un- Firms do not always have an opportu- dertake investments that individuals or nity to delay investments. For example, other firms cannot undertake. Most im- there can be occasions in which strategic portant, these options to invest are valu- considerations make it imperative for a able. Indeed, for most firms, a substantial firm to invest quickly and thereby part of their market value is attributable preempt investment by existing or po- to their options to invest and grow in This content downloaded from 18.7.29.240 on Wed, 5 Nov 2014 09:36:58 AM All use subject to JSTOR Terms and Conditions 1112 Journal of Economic Literature, Vol. XXIX (September 1991) the future, as opposed to the capital they This paper has several objectives. already have in place. (For discussions First, I will review some basic models of growth options as sources of firm of irreversible investment to illustrate value, see Stewart Myers 1977; W. Carl the option-like characteristics of invest- Kester 1984; and my 1988 article.) ment opportunities, and to show how op- When a firm makes an irreversible in- timal investment rules can be obtained vestment expenditure, it exercises, or from methods of option pricing, or alter- "kills," its option to invest. It gives up natively from dynamic programming. Be- the possibility of waiting for new informa- sides demonstrating a methodology that tion to arrive that might affect the desir- can be used to solve a class of investment ability or timing of the expenditure; it problems, this will show how the result- cannot disinvest should market condi- ing investment rules depend on various tions change adversely. This lost option parameters that come from the market value is an opportunity cost that must environment. be included as part of the cost of the A second objective is to survey briefly investment. As a result, the NPV rule some recent applications of this method- "Invest when the value of a unit of capital ology to a variety of investment prob- is at least as large as its purchase and lems, and to the analysis of firm and in- installation cost" must be modified. The dustry behavior. Examples will include value of the unit must exceed the pur- the effects of sunk costs of entry, exit, chase and installation cost, by an amount and temporary shutdowns and restartups equal to the value of keeping the invest- on investment and output decisions, the ment option alive. implications of construction time (and the Recent studies have shown that this option to abandon construction) for the opportunity cost of investing can be value of a project, and the determinants large, and investment rules that ignore of a firm's choice of capacity.
Load more

Recommended publications
  • Calculating the Value Function When the Bellman Equation Cannot Be
    Journal of Economic Dynamics & Control 28 (2004) 1437–1460 www.elsevier.com/locate/econbase Investment under uncertainty: calculating the value function when the Bellman equation cannot besolvedanalytically Thomas Dangla, Franz Wirlb;∗ aVienna University of Technology, Theresianumgasse 27, A-1040 Vienna, Austria bUniversity of Vienna, Brunnerstr.˝ 72, A-1210 Vienna, Austria Abstract The treatment of real investments parallel to ÿnancial options if an investment is irreversible and facing uncertainty is a major insight of modern investment theory well expounded in the book Investment under Uncertainty by Dixit and Pindyck. Thepurposeof this paperis to draw attention to the fact that many problems in managerial decision making imply Bellman equations that cannot be solved analytically and to the di6culties that arise when approximating therequiredvaluefunction numerically.This is so becausethevaluefunctionis thesaddlepoint path from the entire family of solution curves satisfying the di7erential equation. The paper uses a simple machine replacement and maintenance framework to highlight these di6culties (for shooting as well as for ÿnite di7erence methods). On the constructive side, the paper suggests and tests a properly modiÿed projection algorithm extending the proposal in Judd (J. Econ. Theory 58 (1992) 410). This fast algorithm proves amendable to economic and stochastic control problems(which onecannot say of theÿnitedi7erencemethod). ? 2003 Elsevier B.V. All rights reserved. JEL classiÿcation: D81; C61; E22 Keywords: Stochastic optimal control; Investment under uncertainty; Projection method; Maintenance 1. Introduction Without doubt, thebook of Dixit and Pindyck (1994) is oneof themost stimulating books concerning management sciences. The basic message of this excellent book is the following: Considering investments characterized by uncertainty (either with ∗ Corresponding author.
    [Show full text]
  • From Single-Agent to Multi-Agent Reinforcement Learning: Foundational Concepts and Methods Learning Theory Course
    From Single-Agent to Multi-Agent Reinforcement Learning: Foundational Concepts and Methods Learning Theory Course Gon¸caloNeto Instituto de Sistemas e Rob´otica Instituto Superior T´ecnico May 2005 Abstract Interest in robotic and software agents has increased a lot in the last decades. They allow us to do tasks that we would hardly accomplish otherwise. Par- ticularly, multi-agent systems motivate distributed solutions that can be cheaper and more efficient than centralized single-agent ones. In this context, reinforcement learning provides a way for agents to com- pute optimal ways of performing the required tasks, with just a small in- struction indicating if the task was or was not accomplished. Learning in multi-agent systems, however, poses the problem of non- stationarity due to interactions with other agents. In fact, the RL methods for the single agent domain assume stationarity of the environment and cannot be applied directly. This work is divided in two main parts. In the first one, the reinforcement learning framework for single-agent domains is analyzed and some classical solutions presented, based on Markov decision processes. In the second part, the multi-agent domain is analyzed, borrowing tools from game theory, namely stochastic games, and the most significant work on learning optimal decisions for this type of systems is presented. ii Contents Abstract 1 1 Introduction 2 2 Single-Agent Framework 5 2.1 Markov Decision Processes . 6 2.2 Dynamic Programming . 10 2.2.1 Value Iteration . 11 2.2.2 Policy Iteration . 12 2.2.3 Generalized Policy Iteration . 13 2.3 Learning with Model-free methods .
    [Show full text]
  • Dynamic Programming for Dummies, Parts I & II
    DYNAMIC PROGRAMMING FOR DUMMIES Parts I & II Gonçalo L. Fonseca [email protected] Contents: Part I (1) Some Basic Intuition in Finite Horizons (a) Optimal Control vs. Dynamic Programming (b) The Finite Case: Value Functions and the Euler Equation (c) The Recursive Solution (i) Example No.1 - Consumption-Savings Decisions (ii) Example No.2 - Investment with Adjustment Costs (iii) Example No. 3 - Habit Formation (2) The Infinite Case: Bellman's Equation (a) Some Basic Intuition (b) Why does Bellman's Equation Exist? (c) Euler Once Again (d) Setting up the Bellman's: Some Examples (i) Labor Supply/Household Production model (ii) Investment with Adjustment Costs (3) Solving Bellman's Equation for Policy Functions (a) Guess and Verify Method: The Idea (i) Guessing the Policy Function u = h(x) (ii) Guessing the Value Function V(x). (b) Guess and Verify Method: Two Examples (i) One Example: Log Return and Cobb-Douglas Transition (ii) The Other Example: Quadratic Return and Linear Transition Part II (4) Stochastic Dynamic Programming (a) Some Basics (b) Some Examples: (i) Consumption with Uncertainty (ii) Asset Prices (iii) Search Models (5) Mathematical Appendices (a) Principle of Optimality (b) Existence of Bellman's Equation 2 Texts There are actually not many books on dynamic programming methods in economics. The following are standard references: Stokey, N.L. and Lucas, R.E. (1989) Recursive Methods in Economic Dynamics. (Harvard University Press) Sargent, T.J. (1987) Dynamic Macroeconomic Theory (Harvard University Press) Sargent, T.J. (1997) Recursive Macroeconomic Theory (unpublished, but on Sargent's website at http://riffle.stanford.edu) Stokey and Lucas does more the formal "mathy" part of it, with few worked-through applications.
    [Show full text]
  • The Uncertainty Bellman Equation and Exploration
    The Uncertainty Bellman Equation and Exploration Brendan O’Donoghue 1 Ian Osband 1 Remi Munos 1 Volodymyr Mnih 1 Abstract tions that maximize rewards given its current knowledge? We consider the exploration/exploitation prob- Separating estimation and control in RL via ‘greedy’ algo- lem in reinforcement learning. For exploitation, rithms can lead to premature and suboptimal exploitation. it is well known that the Bellman equation con- To offset this, the majority of practical implementations in- nects the value at any time-step to the expected troduce some random noise or dithering into their action value at subsequent time-steps. In this paper we selection (such as -greedy). These algorithms will even- consider a similar uncertainty Bellman equation tually explore every reachable state and action infinitely (UBE), which connects the uncertainty at any often, but can take exponentially long to learn the opti- time-step to the expected uncertainties at subse- mal policy (Kakade, 2003). By contrast, for any set of quent time-steps, thereby extending the potential prior beliefs the optimal exploration policy can be com- exploratory benefit of a policy beyond individual puted directly by dynamic programming in the Bayesian time-steps. We prove that the unique fixed point belief space. However, this approach can be computation- of the UBE yields an upper bound on the vari- ally intractable for even very small problems (Guez et al., ance of the posterior distribution of the Q-values 2012) while direct computational approximations can fail induced by any policy. This bound can be much spectacularly badly (Munos, 2014). tighter than traditional count-based bonuses that For this reason, most provably-efficient approaches to re- compound standard deviation rather than vari- inforcement learning rely upon the optimism in the face of ance.
    [Show full text]
  • Macroeconomics: a Dynamic General Equilibrium Approach
    Macroeconomics: A Dynamic General Equilibrium Approach Mausumi Das Lecture Notes, DSE Jan 29-Feb 19, 2019 Das (Lecture Notes, DSE) DGE Approach Jan29-Feb19,2019 1/113 Modern Macroeconomics: the Dynamic General Equilibrium (DGE) Approach Modern macroeconomics is based on a dynamic general equilibrium approach which postulates that Economic agents are continuously optimizing/re-optimizing subject to their constraints and subject to their information set. They optimize not only over their current choice variables but also the choices that would be realized in future. All agents have rational expectations: thus their ex ante optimal future choices would ex post turn out to be less than optimal if and only if their information set is incomplete and/or there are some random elements in the economy which cannot be anticipated perfectly. The agents are atomistic is the sense that they treat the market factors as exogenous in their optimization exercise. The optimal choices of all agents are then mediated through the markets to produce an equilibrium outcome for the macroeconomy (which, by construction, is also consistent with the optimal choice of each agent). Das (Lecture Notes, DSE) DGE Approach Jan29-Feb19,2019 2/113 Modern Macroeconomics: DGE Approach (Contd.) This approach is ‘dynamic’because agents are making choices over variables that relate to both present and future. This approach is ‘equilibrium’because the outcome for the macro-economy is the aggregation of individuals’equilibrium (optimal) behaviour. This approach is ‘general equilibrium’because it simultaneously takes into account the optimal behaviour of diiferent types of agents in different markets and ensures that all markets clear.
    [Show full text]
  • Economics 2010C: Lecture 1 Introduction to Dynamic Programming
    Economics 2010c: Lecture 1 Introduction to Dynamic Programming David Laibson 9/02/2014 Outline of my half-semester course: 1. Discrete time methods (Bellman Equation, Contraction Mapping Theorem, and Blackwell’s Sufficient Conditions, Numerical methods) Applications to growth, search, consumption, asset pricing • 2. Continuous time methods (Bellman Equation, Brownian Motion, Ito Process, and Ito’s Lemma) Applications to search, consumption, price-setting, investment, indus- • trial organization, asset-pricing Outline of today’s lecture: 1. Introduction to dynamic programming 2. The Bellman Equation 3. Three ways to solve the Bellman Equation 4. Application: Search and stopping problem 1 Introduction to dynamic programming. Course emphasizes methodological techniques and illustrates them through • applications. We start with discrete-time dynamic optimization. • Is optimization a ridiculous model of human behavior? Why or why not? • Today we’ll start with an -horizon stationary problem: • ∞ The Sequence Problem (cf. Stokey and Lucas) Notation: is the state vector at date (+1) is the flow payoff at date ( is ‘stationary’) is the exponential discount function is referred to as the exponential discount factor The discount rate is the rate of decline of the discount function, so ln = ≡− − h i Note that exp( )= and exp( )= − − Definition of Sequence Problem: Find () such that ∞ (0)= sup (+1) +1 =0 { }∞=0 X subject to +1 Γ() with 0 given. ∈ Remark 1.1 When I omit time subscripts, this implies that an equation holds for all relevant values of . In the statement above, +1 Γ() implies, ∈ +1 Γ() for all =0 1 2 ∈ Example 1.1 Optimal growth with log utility and Cobb-Douglas technology: ∞ sup ln() =0 { }∞=0 X subject to the constraints, 0+ +1 = and 0 given.
    [Show full text]
  • Graduate Macro Theory II: Notes on Investment
    Graduate Macro Theory II: Notes on Investment Eric Sims University of Notre Dame Spring 2011 1 Introduction These notes introduce and discuss modern theories of firm investment. While much of this is done as a decision rule problem of the firm, it is easily incorporated into a general equilibrium structure. 2 Tobin's Q Jim Tobin (1969) developed an intuitive and celebrated theory of investment. He reasoned that if the market value of physical capital of a firm exceeded its replacement cost, then capital has more value \in the firm” (the numerator) than outside the firm (the denominator). Formally, Tobin's Q is defined as: Market Value of Firm Capital Q = (1) Replacement Cost of Capital Tobin reasoned that firms should accumulate more capital when Q > 1 and should draw down their capital stock when Q < 1. That is, net investment in physical capital should depend on where Q is in relation to one. How would one measure Q in the data? Typically this is done by using the value of the stock market in the numerator and data on corporate net worth in the denominator { here net worth is defined as the replacement cost of capital goods. For example, suppose that a company is in the business of bulldozing. It owns a bunch of bulldozers. Suppose that it owns 100 bulldozers, which cost 100 each to replace. This means its replacement cost is 10; 000. Suppose it has 1000 outstanding shares of stock, valued at 15 per share. The market value of the firm is thus 15; 000 and the replacement cost is 10; 000, so Tobin's Q is 1.5.
    [Show full text]
  • Chapter 3 the Representative Agent Model
    Chapter 3 The representative agent model 3.1 Optimal growth In this course we’re looking at three types of model: 1. Descriptive growth model (Solow model): mechanical, shows the im- plications of a given fixed savings rate. 2. Optimal growth model (Ramsey model): pick the savings rate that maximizes some social planner’s problem. 3. Equilibrium growth model: specify complete economic environment, find equilibrium prices and quantities. Two basic demographies - rep- resentative agent (RA) and overlapping generations (OLG). 3.1.1 The optimal growth model in discrete time Time and demography Time is discrete. One representative consumer with infinite lifetime. Social planner directly allocates resources to maximize consumer’s lifetime expected utility. 1 2 CHAPTER 3. THE REPRESENTATIVE AGENT MODEL Consumer The consumer has (expected) utility function: ∞ X t U = E0 β u(ct) (3.1) t=0 where β ∈ (0, 1) and the flow utility or one-period utility function is strictly increasing (u0(c) > 0) and strictly concave (u00(c) < 0). In addition we will impose an Inada-like condition: lim u0(c) = ∞ (3.2) c→0 The social planner picks a nonnegative sequence (or more generally a stochas- tic process) for ct that maximizes expected utility subject to a set of con- straints. The technological constraints are identical to those in the Solow model. Specifically, the economy starts out with k0 units of capital at time zero. Output is produced using labor and capital with the standard neoclassical production function. Labor and technology are assumed to be constant, and are normalized to one. yt = F (kt,Lt) = F (kt, 1) (3.3) Technology and labor force growth can be accommodated in this model by applying the previous trick of restating everything in terms of “effective la- bor.” Output can be freely converted to consumption ct, or capital investment xt, as needed: ct + xt ≤ yt (3.4) Capital accumulation is as in the Solow model: kt+1 = (1 − δ)kt + xt (3.5) Finally, the capital stock cannot fall below zero.
    [Show full text]
  • Q-Function Learning Methods
    Q-Function Learning Methods February 15, 2017 Value Functions I Definitions (review): π 2 Q (s; a) = Eπ r0 + γr1 + γ r2 + ::: j s0 = s; a0 = a Called Q-function or state-action-value function π 2 V (s) = Eπ r0 + γr1 + γ r2 + ::: j s0 = s π = Ea∼π [Q (s; a)] Called state-value function Aπ(s; a) = Qπ(s; a) − V π(s) Called advantage function Bellman Equations for Qπ π I Bellman equation for Q π π Q (s0; a0) = Es1∼P(s1 j s0;a0) [r0 + γV (s1)] π = Es1∼P(s1 j s0;a0) [r0 + γEa1∼π [Q (s1; a1)]] π I We can write out Q with k-step empirical returns π π Q (s0; a0) = Es1;a1 j s0;a0 [r0 + γV (s1; a1)] 2 π = Es1;a1;s2;a2 j s0;a0 r0 + γr1 + γ Q (s2; a2) h k−1 k π i = Es1;a1:::;sk ;ak j s0;a0 r0 + γr1 + ··· + γ rk−1 + γ Q (sk ; ak ) Bellman Backups I From previous slide: π π Q (s0; a0) = Es1∼P(s1 j s0;a0) [r0 + γEa1∼π [Q (s1; a1)]] I Define the Bellman backup operator (operating on Q-functions) as follows π [T Q](s0; a0) = Es1∼P(s1 j s0;a0) [r0 + γEa1∼π [Q(s1; a1)]] π I Then Q is a fixed point of this operator T πQπ = Qπ π I Furthermore, if we apply T repeatedly to any initial Q, the series converges to Qπ Q; T πQ; (T π)2Q; (T π)3Q; ···! Qπ Introducing Q∗ ∗ I Let π denote an optimal policy ∗ π∗ ∗ π I Define Q = Q , which satisfies Q (s; a) = maxπ Q (s; a) ∗ ∗ ∗ I π satisfies π (s) = arg maxa Q (s; a) I Thus, Bellman equation π π Q (s0; a0) = Es1∼P(s1 j s0;a0) [r0 + γEa1∼π [Q (s1; a1)]] becomes ∗ ∗ Q (s0; a0) = Es1∼P(s1 j s0;a0) r0 + γ max Q (s1; a1) a1 I Another definition ∗ ∗ Q (s0; a0) = Es1 [r + γV (s1)] Bellman Operator for Q∗ I Define a corresponding
    [Show full text]
  • Optimal and Autonomous Control Using Reinforcement Learning: a Survey
    2042 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 29, NO. 6, JUNE 2018 Optimal and Autonomous Control Using Reinforcement Learning: A Survey Bahare Kiumarsi , Member, IEEE , Kyriakos G. Vamvoudakis, Senior Member, IEEE , Hamidreza Modares , Member, IEEE , and Frank L. Lewis, Fellow, IEEE Abstract — This paper reviews the current state of the art on unsupervised, or reinforcement, depending on the amount and reinforcement learning (RL)-based feedback control solutions quality of feedback about the system or task. In supervised to optimal regulation and tracking of single and multiagent learning, the feedback information provided to learning algo- systems. Existing RL solutions to both optimal H2 and H∞ control problems, as well as graphical games, will be reviewed. rithms is a labeled training data set, and the objective is RL methods learn the solution to optimal control and game to build the system model representing the learned relation problems online and using measured data along the system between the input, output, and system parameters. In unsu- trajectories. We discuss Q-learning and the integral RL algorithm pervised learning, no feedback information is provided to the as core algorithms for discrete-time (DT) and continuous-time algorithm and the objective is to classify the sample sets to (CT) systems, respectively. Moreover, we discuss a new direction of off-policy RL for both CT and DT systems. Finally, we review different groups based on the similarity between the input sam- several applications. ples. Finally, reinforcement learning (RL) is a goal-oriented Index Terms — Autonomy, data-based optimization, learning tool wherein the agent or decision maker learns a reinforcement learning (RL).
    [Show full text]
  • Insurance and Taxation Over the Life Cycle
    Insurance and Taxation over the Life Cycle EMMANUEL FARHI IVÁN WERNING Harvard University MIT August 2012 Abstract We consider a dynamic Mirrlees economy in a life cycle context and study the op- timal insurance arrangement. Individual productivity evolves as a Markov process and is private information. We use a first order approach in discrete and continuous time and obtain novel theoretical and numerical results. Our main contribution is a formula describing the dynamics for the labor-income tax rate. When productivity is an AR(1) our formula resembles an AR(1) with a trend where: (i) the auto-regressive coefficient equals that of productivity; (ii) the trend term equals the covariance pro- ductivity with consumption growth divided by the Frisch elasticity of labor; and (iii) the innovations in the tax rate are the negative of consumption growth. The last prop- erty implies a form of short-run regressivity. Our simulations illustrate these results and deliver some novel insights. The average labor tax rises from 0% to 37% over 40 years, while the average tax on savings falls from 12% to 0% at retirement. We com- pare the second best solution to simple history independent tax systems, calibrated to mimic these average tax rates. We find that age dependent taxes capture a sizable fraction of the welfare gains. In this way, our theoretical results provide insights into simple tax systems. 1 Introduction To a twenty five year old entering the labor market, the landscape must feel full of un- certainties. Will they land a good job relatively quickly or will they initially bounce from one job to another in search of a good match? What opportunities for on-the-job training and other forms of skill accumulation be they find? How well will they take advantage of these opportunities? Just how good are they? How high will they rise? Will they ad- vance steadily within a firm or industry, or be laid off and have to reinvent themselves 1 elsewhere? For all these reasons, young workers must find it challenging to predict how much they will be making at, say, age fifty.
    [Show full text]
  • Bellman Equation
    Harvard University, Neurobiology 101hfm. Fundamentals in Computational Neuroscience Spring term 2014/15 Lecture 8 – Reinforcement Learning III: Bellman equation Alexander Mathis, Ashesh Dhawale April 7st, 2015, due date April, 15th, 2015 In class we have treated various aspects of Reinforcement learning (RL) so far, in particular the Rescola-Wagner model, greedy and softmax policies, and temporal difference (TD) learning. Here we provide a more principled, overarching view of RL with a focus on the mathematics rather than neurobiology. We largely follow the exposition of the classical book by Sutton and Barto called “Reinforcement Learning”.1 1 The Reinforcement learning problem RL studies how an agent is learning from interactions with an environment in order to achieve a goal. At each time t, which evolves in discrete steps, the agent perceives some state st 2 S, where S is the (discrete) set of possible states. and performs an action at 2 A(st), where A(st) is the set of actions possible in state st. In the next time point the agent receives a scalar reward rt+1 2 R and senses state st+1. This agent-environment interface is shown in Figure 1. Figure 1: The agent-environment interaction in RL. Figure from Sutton/Barto. At each time point the agent acts according to a probabilistic ’rule’, which maps states S onto actions A and is called policy π. In order to emphasize those dependencies one can write πt(s; a). The agent’s goal in RL is to maximize its total reward collected in the environment, called the return Rt.
    [Show full text]