Y. Freund, M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, and R. E. Shapire. Efficient Algorithms for Learning to Play Repeated Games Against Computationally Bounded Adversaries. In 36th Annual Symposium on Foundations of Computer Science (FOCS'95), pages 332--343, Los Alamitos, CA, 1995. IEEE Computer Society Press.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....because no reasoning is conducted about the interaction itself by the agents. Instead, mathematical analysis of decision situations (as in the area of mechanism design ( 4] provides an overview) or data driven adaptation of learning algorithms (especially in work on game learning, cf. e.g. [5]) is employed to yield the desired behaviour. In our work, we followed a different approach. We investigated the possibilities of learning how to interact effectively in repeated n player games on the grounds of developing a common sense (one could even say naive ) understanding of the ....

Y. Freund, D. Ron, M. Kearns, R. Rubinfeld, Y. Mansour, and R. E. Schapire. Efficient Algorithms for Learning to Play Repeated Games Against Computationally Bounded Adversaries. Proceedings of the 36th Annual Symposium on Foundations of Computer Science, 1995.

....because no reasoning is conducted about the interaction itself by the agents. Instead, mathematical analysis of decision situations (as in the area of mechanism design ( 4] provides an overview) or data driven adaptation of learning algorithms (especially in work on game learning, cf. e.g. [5]) is employed to yield the desired behaviour. In our work, we followed a different approach. We investigated the possibilities of learning how to interact effectively in repeated n player games on the grounds of developing a common sense (one could even say naive ) understanding of the ....

Y. Freund, D. Ron, M. Kearns, R. Rubinfeld, Y. Mansour, and R. E. Schapire. Efficient Algorithms for Learning to Play Repeated Games Against Computationally Bounded Adversaries. Proceedings of the 36th Annual Symposium on Foundations of Computer Science, 1995.

....for value functions. These are not useful for complex domains in which the number of states is vast, but there have been promising recent results using certain kinds of value function approximators [10] Several recent papers have discussed adaptive strategies for simple repeated games [13, 8]. The goal of such strategies is typically to learn enough about a particular opponent in early games to do well against it in later games. These concerns are largely orthogonal to our goal of learning strategies that are robust against a large space of opponents. 2.4.2 Experimental A motivating ....

Freund, Y., M. Kearns et al. (1995) Efficient Algorithms for Learning to Play Repeated Games Against Computationally Bounded Adversaries. FOCS 36.

....according to the given distribution. Second, find the best response automaton against the probabilistic product automaton. The problem of finding the best response against a probabilistic action automaton (PAA) is equivalent to the best response problem against a deterministic automaton [8]. The above interpretation still assumes that the agent does not modify its opponent model throughout the game path while computing the utility, and is therefore not suitable for the belief revision framework. The second interpretation considers the opponent s strategy to be one of the models ....

Y. Freund, M. Kearns, Y. Mansour, D. Ron, R. Rubinfeled, and R. E. Schapire. Efficient algorithms for learning to play repeated games against computationally bounded adversaries. In Proceedings of the Annual Symposium on the Foundations of Computer Science, pages 332--341, 1995.

....for infinite games and is justified by the fact that learning and uniform computation turn out to be incomparable. We also consider the concept of counter strategy learning for closed recursive games. The idea of counter strategy learning is taken from Fortnow and Whang [6] and Freund et al. [7], who have studied this concept in the framework of repeated matrix games. 2 Notation and Definitions The natural numbers are denoted by . A is the characteristic function of A . We are using an acceptable programming system 0 ; 1 ; the function computed by the e th program ....

....a 2 f0; 1g such that a b 2 G. From the above reductions between games and trees we get al..so the following equivalence: Theorem 29 Play learning for closed recursive games is equivalent to branch enumerating for trees. 5. 4 Counter Strategy Learning Fortnow and Whang [6] and Freund et al. [7] have studied a notion of counter strategy learning in repeated matrix games. In these games there are always optimal strategies for the two players. But when the computational resources of the adversary are limited he may not be able to execute the optimal strategy. Is it possible to learn for ....

Y. Freund, M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, and R. E. Schapire. Efficient algorithms for learning to play repeated games against computationally bounded adversaries. In 36th Annual Symposium on Foundations of Computer Science, pages 332--341, Milwaukee, Wisconsin, 23--25 Nov. 1995. IEEE.

....for value functions. These are not useful for complex domains in which the number of states is vast, but there have been promising recent results using certain kinds of value function approximators [9] Several recent papers have discussed adaptive strategies for simple repeated games [11, 5]. The goal of such strategies is typically to learn enough about a particular opponent in early games to do well against it in later games. These concerns are largely orthogonal to our goal of learning strategies that are robust against a large space of opponents. 3 Competitive Algorithm ....

Freund, Y., M. Kearns et al. (1995) Efficient Algorithms for Learning to Play Repeated Games Against Computationally Bounded Adversaries. FOCS 36.

No context found.

Y. Freund, M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, and R. E. Schapire. Efficient algorithms for learning to play repeated games against computationally bounded adversaries. In Proceedings of the Thirty Sixth Annual Symposium on Foundations of Computer Science, pages 332--341, 1995.

No context found.

Freud Y., Kearns M., Mansour Y., Ron D., Rubinfeld R., and Schapire R.E. (1995). Efficient Algorithms for Learning to Play Repeated Games Against Computationally Bounded Adversaries. In Proceedings of the 36th Annual Symposium on Foundations of Computer Science.

....actually learns the topology of the underlying graph. Their algorithms (with the exception of one, for permutation automata) rely on a teacher. The teacher supplies counterexamples to the robot s hypotheses. Variants of this problem that do not rely on a teacher are studied in the following works [14, 18, 28, 17]. We note that Dean et al. 14] apply a cycling technique related to ours, but for different purposes. Exploring and navigating in geometric environments is studied extensively. A sample of papers includes [5, 23, 15, 11, 6, 10, 8, 19, 2] 2 Preliminaries Let G = V;E) be the unknown directed ....

Y. Freund, M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, and R. E. Schapire. Efficient algorithms for learning to play repeated games against computationally bounded adversaries. In Proceedings of the Thirty Sixth Annual Symposium on Foundations of Computer Science, pages 332--341, 1995.

....actually learns the topology of the underlying graph. Their algorithms (with the exception of one, for permutation automata) rely on a teacher. The teacher supplies counterexamples to the robot s hypotheses. Variants of this problem that do not rely on a teacher are studied in the following works [14, 18, 27, 17]. We note that Dean et al. 14] apply a cycling technique related to ours for different purposes. Exploring and navigating in geometric environments is studied extensively. A sample of papers includes [5, 22, 15, 11, 6, 10, 8, 19, 2] 2 Preliminaries Let G = V; E) be the unknown directed graph ....

Y. Freund, M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, and R. E. Schapire. Efficient algorithms for learning to play repeated games against computationally bounded adversaries. In Proceedings of the Thirty Sixth Annual Symposium on Foundations of Computer Science, pages 332--341, 1995.

....However, when the learner does not have means of a reset, and thus performs a single walk on M , we know of no natural notion of approximately correct learning. In particular, it is not clear what type of approximation suffices for the game theoretical scenario. In recent work of Freund et al. [11] our results have been improved as follows. Freund et al. consider the problem of learning probabilistic output automata. These are finite automata whose transition function is deterministic, but whose output function is probabilistic. Namely, for any given string, whenever performing the walk ....

....of flipping a coin with a bias that depends on the state reached. In the case when the biases at each state are either j or 1 Gamma j for some 0 j 1=2, this is essentially the problem of learning deterministic automata in the presence of noise, for which we give an algorithm in this paper. In [11], a learning algorithm is given that runs in time polynomial in the cover time of the target automaton, with no restrictions on the biases at each state. Other Related Work Several researchers have considered the problem of learning DFAs in the limit. In this setting the learner is presented ....

Y. Freund, M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, and R. Schapire. Efficient algorithms for learning to play repeated games against computationally bounded adversaries. To appear in Proceedings of the Thirty Sixth Annual Symposium on Foundations of Computer Science, 1995.

....to a PAC learning algorithm with membership queries, DFAs are efficiently learnable in this model. However, when the learner does not have means of a reset, and thus performs a single walk on M , we know of no natural notion of approximately correct learning. In recent work of Freund et al. [13] our results have been improved as follows. Freund et al. consider the problem of learning probabilistic output automata. These are finite automata whose transition function is deterministic, but whose output function is probabilistic. Namely, for any given string, whenever performing the walk ....

....of flipping a coin with a bias that depends on the state reached. In the case when the biases at each state are either j or 1 Gamma j for some 0 j 1=2, this is essentially the problem of learning deterministic automata in the presence of noise, for which we give an algorithm in this paper. In [13], a learning algorithm is given that runs in time polynomial in the cover time of the target automaton, with no restrictions on the biases at each state. Repeated games against computationally bounded opponents Another motivation for this work is the game theoretical problem of finding an optimal ....

Y. Freund, M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, and R. Schapire. Efficient algorithms for learning to play repeated games against computationally bounded adversaries. In Proceedings of the Thirty Seventh Annual Symposium on Foundations of Computer Science, pages 332--341, 1996.

No context found.

Y. Freund, M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, and R. E. Shapire. Efficient Algorithms for Learning to Play Repeated Games Against Computationally Bounded Adversaries. In 36th Annual Symposium on Foundations of Computer Science (FOCS'95), pages 332--343, Los Alamitos, CA, 1995. IEEE Computer Society Press.

No context found.

Yoav Freund, Michael Kearns, Yishay Mansour, Dana Ron, Ronitt Rubinfeled, and Robert E. Schapire. Efficient algorithms for learning to play repeated games against computationally bounded adversaries. In Proceedings of the Annual Symposium on the Foundations of Computer Science, pages 332--341, 1995.

No context found.

Y. Freund, M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, and R. E. Schapire. Efficient algorithms for learning to play repeated games against computationally bounded adversaries. In IEEE Symposium on the Foundations of Computer Science, 1995.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC