Abstract
We use dynamics of measures, i.e. iteration of the operators from measurable space to space of probabilistic measures on this space, to model and prove properties of random search algorithms. Specifically using this technique in the context of Game Theory we show that stochastic better response dynamics, where players in the potential game perform their moves independently choosing the random strategy improving their outcome, converges in stochastic sense to playing strategies near equilibrium.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
Opposite direction, where one has simple method of sampling for A(x) but sampling X is difficult, is also used in practice. When \(\mu _{0}\) is stationary distribution of Markov process given by A, famous Metropolis algorithm is such an example.
- 2.
To some extent one can look at the stochastic search method as having significant advantage over the deterministic one, because deterministic method relies on some internal structure of the X. If function we are maximizing is incompatible with this internal structure, deterministic search may be completely lost. A method of sampling assumed is in this case a way to escape from this pitfall. On the other hand, ability to sample requires some source of randomness - using of pseudo-random variables for sampling makes the method de facto deterministic. Intuition when such a method cannot work efficiently comes from Algorithmic Information Theory. In purely computable setting there is no randomness. This is an interesting philosophical aspect of random search.
- 3.
To some extent it can model situation where players are continuously playing their strategies and making decisions about change of the strategy in the moments that are distributed according to a Poisson process. Probability of choice of the same moments for some players is then 0, so we can concentrate only on the order of changes.
References
Giry, M.: A categorical approach to probability theory. In: Banaschewski, B. (ed.) Categorical Aspects of Topology and Analysis. LNM, vol. 915, pp. 68–85. Springer, Heidelberg (1982). https://doi.org/10.1007/BFb0092872
Gutjahr, W.J.: Stochastic search in metaheuristics. In: Gendreau, M., Potvin, J.Y. (eds.) Handbook of Metaheuristics. ISOR, vol. 146, pp. 573–597. Springer, Boston (2010). https://doi.org/10.1007/978-1-4419-1665-5_19
MacKenzie, A.B., DaSilva, L.A.: Game Theory for Wireless Engineers. Synthesis Lectures on Communications. Morgan and Claypool Publishers, San Rafael (2006)
Ombach, J., Tarłowski, D.: Nonautonomous stochastic search in global optimization. J. Nonlinear Sci. 22(2), 169–185 (2012)
Solis, F.J., Wets, R.J.B.: Minimization by random search techniques. Math. Oper. Res. 6(1), 19–30 (1981)
Tarłowski, D.: Nonautonomous dynamical systems in stochastic global optimization. Ph.D. thesis, Department of Mathematics, Jagiellonian University (2014)
Young, H.P.: Learning by trial and error. Games Econ. Behav. 65(2), 626–643 (2009)
Łojasiewicz, S.: An Introduction to the Theory of Real Functions. Wiley, Hoboken (1988)
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
A Appendix
A Appendix
Following is the proof of measurability of the function f defined in the proof of Theorem 8. We will infer it from two lemmas.
Lemma 1
Let P be the finite set, and for each \(i \in P\) \(S_{i}\) be a compact metric space with metric \(d_{i}\). For each i by \(\sigma _{i}\) we will denote \(\sigma \text {-algebra}\) of Borel sets and \(\mu _{i}\) be a regular probabilistic measure on \(S_{i}\). Let \(S=\prod \limits _{i \in P}\) be a metric space with metric d:
which gives a product topology in S. We denote by \(\sigma \) the \(\sigma \text {-algebra}\) of Borel set in S, and take measure \(\mu =\bigotimes \limits _{i \in P}\mu _{i}\). Let \(V:S \rightarrow \mathbb {R}\) be continuous, \(\alpha \ge 0\) and function \(g_{i}: S \rightarrow \mathbb {R}\) be defined as follows:
where \(U_{t}=V^{-1}((t,\infty ))\). Under this assumptions we have: \(g_{i}\) is lower semi-continuous
Proof
We will show, that \(g^{-1}_{i}((\rho , \infty ))\) for \(\rho \in \mathbb {R}\) is open. Let’s assume that \(g^{-1}_{i}((\rho , \infty )) \ne \emptyset \) (if it is empty, it is open trivially) and take \(s \in g^{-1}_{i}((\rho , \infty ))\). From definition we have:
Since \(\mu _{i}\) is regular, there is a compact set K such that \(K \subset \iota ^{-1}_{i,s}(U_{V(s)+\alpha })\) and \(\mu _{i} > \rho \) Obviously, \(\iota _{i,s}(K)\) is compact so V takes it’s extrema on K. Thus, there is \(\epsilon > 0\) such that, for each \(x \in \iota _{i,s}(K)\) we have \(V(x) > V(s)+\alpha +\epsilon \) For each \(x \in \iota _{i,s}(K)\) then, there is open open neighborhood \(O_{x}\) of x such that: \(O_{x} = \prod \limits _{j \in P}O^{j}_{x}\), where \(O^{j}_{x}\) is neighborhood of \(x_{j}\) in \(S_{j}\) and for each \(y \in O_{x}\) we have \(V(y) > V(s)+\alpha +\frac{\epsilon }{2}\) From the compactness, we can over \(\iota _{i,s}(K)\) by finite number of these neighborhoods:
where last equality is introduced to simplify further referencing these sets. Now let \(O_{s} = V^{-1}((-\infty , V(s)+\frac{\epsilon }{2}))\). It is open neighborhood of s in S. Finally let’s construct another open neighborhood of s following way:
where \(O^{-i}_{j} = \prod \limits _{l \in (P-\{i\})}O^{l}_{j}\)
We will show, that \(O \subset g^{-1}_{i}((\rho , \infty ))\).
Let’s take \(s' \in O\). For each \(x \in K\) we have: \(\iota _{i, s'}(x) = (x, s'_{-i})\) where \(s'_{-i} \in \bigcap \limits _{j=0, \ldots , n}O^{-i}_{j}\), thus belongs to all \(O^{-i}_{j}\). Since \(O_{j}\) was cover of \(\iota _{i,s}(K)\), there is specific j, such that \((x,s_{-i}) \in O_{j}\). So, also \((x, s'_{-i}) \in O_{j}\). Now, we have:
thus \(K \subset \iota ^{-1}_{i,s'}(U_{V(s'}+\alpha \) and since \(\mu _{i}(K) > \rho \) we have \(O \subset g^{-1}_{i}((\rho , \infty ))\). This completes the proof.
Lemma 2
Let (X, d) be a metric space, \(f:X\rightarrow \mathbb {R}\) and \(g:X\rightarrow \mathbb {R}\) be two lower semi-continuous functions such that, for each \(x \in M\) we have \(0 \le g(x) \le f(x) \le M\). Let \(\phi :X\rightarrow \mathbb {R}\) be defined as
then \(\phi \) is measurable in the sense of the Borel \(\sigma \text {-algebra}\)
Proof
We will show, that \(\phi \) is Baire class 2 function (so measurable with regard to Borel’s \(\sigma \text {-algebra}\). Let’s take sequence \((\epsilon _{n})_{n \in \mathbb {N}}\), such that: \(\epsilon _{n}\underset{n \rightarrow \infty }{\longrightarrow }0\) for each \(n \in \mathbb {N}\) we have \(\epsilon _{n} > \epsilon _{n+1}\).
Obviously functions \(f_{n}=f+\epsilon _{n}\) and \(g_{n} = g+\epsilon _{n}\) are also lower semi-continuous, and what’s more strictly positive (\(>\epsilon _{n}\)). From Baire theorem on lower semi-continuous functions [8] there are pointwise increasing sequence of continuous functions \((f_{n,i})_{i \in \mathbb {N}}\) and \((g_{n,i})_{i \in \mathbb {N}}\) such that pointwise \(f_{n,i}\underset{i \rightarrow \infty }{\longrightarrow }f_{n}\) and \(g_{n,i}\underset{i \rightarrow \infty }{\longrightarrow }g_{n}\). By taking
and
we can assure, that sequences are bounded below from 0 and taking
and
we can assure that \(\tilde{\tilde{g}}_{n,i} \le \tilde{\tilde{f}}_{n,i}\). Now, we have \(\frac{\tilde{\tilde{g}}_{n,i}}{\tilde{\tilde{f}}_{n,i}}\underset{i \rightarrow \infty }{\longrightarrow }\frac{g_{n}}{f_{n}}\) so \(\frac{g_{n}}{f_{n}}\) is of first category as a pointwise limit of continuous functions.
We can also check (by easy calculation), that for each \(x \in X\) \(\frac{g_{n}(x)}{f_{n}(x)} \ge \frac{g_{n+1}(x)}{f_{n+1}(x)}\). Thus \(\frac{g_{n}}{f_{n}}\) is pointwise convergent equal to 1 whenever \(f(x)=g(x)\), so convergent to \(\phi \). Thus \(\phi \) is of second Baire category.
Let us now recall definition of function f from Theorem 8 (we will use notation and assumptions from Sect. 4 and proof of Theorem 8):
We have for each \(i \in P\) that functions \(f_{i}(s)=\tilde{\mathcal {V}}_{i}(s)(U_{i,s})=\mu _{0}^{i}(\iota ^{-1}_{i,s}(U_{V(s)}))\) and \(g_{i}(s)=\tilde{\mathcal {V}}_{i}(s)((U_{V(s)+\epsilon '}))=\mu _{0}^{i}(s)(\iota ^{-1}_{i,s}(U_{V(s)+\epsilon '}))\) both are lower semi-continuous by Lemma 1, and what’s more satisfy conditions of Lemma 2. So, \(\mathcal {V}_{i}(U_{V(s)+\epsilon '})\) is measurable for each i and \(f(s)=A(s)(U_{V(s)+\epsilon '})=\sum \limits _{i \in P}\alpha _{i}\mathcal {V}_{i}(U_{V(s)+\epsilon '})\) is measurable as a linear combination of the measurable functions.
Rights and permissions
Copyright information
© 2019 ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering
About this paper
Cite this paper
Popławski, A. (2019). Games and Random Search. In: Avrachenkov, K., Huang, L., Marden, J., Coupechoux, M., Giovanidis, A. (eds) Game Theory for Networks. GameNets 2019. Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering, vol 277. Springer, Cham. https://doi.org/10.1007/978-3-030-16989-3_3
Download citation
DOI: https://doi.org/10.1007/978-3-030-16989-3_3
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-030-16988-6
Online ISBN: 978-3-030-16989-3
eBook Packages: Computer ScienceComputer Science (R0)