| 1. |
- Fay, Dominik
(författare)
-
Towards Scalable Machine Learning with Privacy Protection
- 2023
-
Licentiatavhandling (övrigt vetenskapligt/konstnärligt)abstract
- The increasing size and complexity of datasets have accelerated the development of machine learning models and exposed the need for more scalable solutions. This thesis explores challenges associated with large-scale machine learning under data privacy constraints. With the growth of machine learning models, traditional privacy methods such as data anonymization are becoming insufficient. Thus, we delve into alternative approaches, such as differential privacy.Our research addresses the following core areas in the context of scalable privacy-preserving machine learning: First, we examine the implications of data dimensionality on privacy for the application of medical image analysis. We extend the classification algorithm Private Aggregation of Teacher Ensembles (PATE) to deal with high-dimensional labels, and demonstrate that dimensionality reduction can be used to improve privacy. Second, we consider the impact of hyperparameter selection on privacy. Here, we propose a novel adaptive technique for hyperparameter selection in differentially gradient-based optimization. Third, we investigate sampling-based solutions to scale differentially private machine learning to dataset with a large number of records. We study the privacy-enhancing properties of importance sampling, highlighting that it can outperform uniform sub-sampling not only in terms of sample efficiency but also in terms of privacy.The three techniques developed in this thesis improve the scalability of machine learning while ensuring robust privacy protection, and aim to offer solutions for the effective and safe application of machine learning in large datasets.
|
|
| 2. |
- Hagelbäck, Johan, 1977-
(författare)
-
A Multi-Agent Potential Field Based Approach for Real-Time Strategy Game Bots
- 2009
-
Licentiatavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Computer games in general and Real-Time Strategy (RTS) games in particular provide a rich challenge for both human- and computer controlled players, often denoted as bots. The player or bot controls a large number of units that have to navigate in partially unknown dynamic worlds to pursue a goal. Navigation in such worlds can be complex and require much computational resources. Typically it is solved by using some sort of path planning algorithm, and a lot of research has been conducted to improve the performance of such algorithms in dynamic worlds. The main goal of this thesis is to investigate an alternative approach for RTS bots based on Artificial Potential Fields, an area originating from robotics. In robotics the technique has successfully been used for navigation in dynamic environments, and we show that it is possible to use Artificial Potential Fields for navigation in an RTS game setting without any need of path planning.In the first three papers we define and demonstrate a methodology for creating multi-agent potential field based bots for an RTS game scenario where two tank armies battle each other. The fourth paper addresses incomplete information about the game world, referred to as the fog of war, and show how Potential Field based bots can handle such environments. The final paper shows how a Potential Field based bot can be evolved to handle a more complex full RTS scenario. It addresses resource gathering, construction of bases, technological development and construction of an army consisting of different types of units.We show that Artificial Potential Fields is a viable option for several RTS game scenarios and that the performance, both in terms of being able to win a game and computational resources used, can match and even surpass those of traditional approaches based on path planning.
|
|
| 5. |
- Johansson, Stefan J.
(författare)
-
Game Theory and Agents
- 1999
-
Licentiatavhandling (övrigt vetenskapligt/konstnärligt)abstract
- A fundamental problem in multi agent systems is conflict resolution. A conflict occurs in general when the agents have to deal with inconsistent goals, such as a demand for shared resources. We investigate how a game theoretic approach may be a helpful and efficient tool in examining a class of conflicts in multi agent systems. We consider the hawk-and-dove game both from an evolutionary and from an iterated point of view. An iterated hawk-and-dove game is not the same as an infinitely repeated evolutionary game because in an iterated game the agents are supposed to know what happened in the previous moves. In an evolutionary game evolutionary stable strategies will be most successful but not necessarily be a unique solution. An iterated game can be modeled as a mixture of a prisoner's dilemma game and a chicken game. These kinds of games are generally supposed to have successful cooperating strategies. We also discuss situations where a chicken game (CG) is a more appropriate model than a prisoner's dilemma (PD) game and describe a simulation of iterated prisoner's dilemma (IPD) and iterated chicken (ICG) games. We study a parameterized class of cooperative games, with these classical games as end cases, and we show that chicken games to a higher extent reward cooperative strategies than defecting strategies. We then introduce generous, even-matched, and greedy strategies as concepts for analyzing games. A two person PD game is described by the four outcomes (C,D), (C,C), (D,C), and (D,D), where the outcome (X,Y) is the probability of that the opponent acts Y, when the own player acts X. In a generous strategy the proportion of (C,D) is larger than that of (D,C), i.e. the probability of facing a defecting agent is larger than the probability of defecting. An even-matched strategy has the (C,D) proportion approximately equal to that of (D,C). A greedy strategy is an inverted generous strategy. The basis of the partition is that it is a zero-sum game given that the sum of the proportions of strategies (C,D) must equal that of (D,C). In a population simulation, we compare the PD game with the CG, given complete as well as partial knowledge of the rules for moves in the other strategies. In a traffic intersection example, we expected a co-operating generous strategy to be successful when the cost for collision was hig h in addition to the presence of uncertainty. The simulation indeed showed that a generous strategy was successful in the CG part, in which agents faced uncertainty about the outcome. If the resulting zero-sum game is changed from a PD game to a CG, or if the noise level is increased, it will favor generous strategies rather than an even-matched or greedy strategies. Four different PD like games were studied by running a round robin as well as a population tournament, using 15 different strategies. The results were analyzed in terms of definitions of generous, even-matched, and greedy strategies. In the round robin, PD favored greedy strategies. CG and coordinate game were favoring generous strategies, and compromise dilemma the unstably even-matched strategy Anti Tit-for-Tat. These results were not surprising because all strategies used were fully dependent on the mutual encounters, not the actual payoff values of the game. A population tournament is a zero-sum game balancing generous and greedy strategies. When strategies disappear, the population will form a new balance between the remaining strategies. A winning strategy in a population tournament has to do well against itself because there will be numerous copies of that strategy. A winning strategy must also be good at resisting invasion from other competing strategies. These restrictions make it natural to look for winning strategies among originally generous or even-matched strategies. For three of the games, this was found true, with original generous strategies being most successful. The most diverging result was that compromise dilemma, despite its close relationship to PD, had two greedy strategies almost entirely dominating the population tournament. In game theory, iterated strategic games are considered harder to analyze than repeated games (for which the theory of mixed strategies is a suitable tool). However, iterated games are in many cases more fit to describe the situation of computerized agents, since it take into account previous moves of the opponents, rather than just assigning each possible action a certain probability. We introduce the notion of characteristic distributions and discuss how it can be used to simplify and structure the analysis of strategies in order to provide a good basis for choosing strategies in games to come and formulate a No free lunch theorem for game theory.
|
|
