Abstract
The problem of making decisions is ubiquitous in life. This problem becomes even more
complex when the decisions should be made sequentially. In fact, the execution of an action
at a given time leads to a change in the environment of the problem, and this change cannot be
predicted with certainty. The aim of a decision-making process is to optimally select actions
in an uncertain environment. To this end, the environment is often modeled as a dynamical
system with multiple states, and the actions are executed so that the system evolves toward
a desirable state.
In this thesis, we proposed a family of stochastic models and algorithms in order to improve
the quality of of the decision-making process. The proposed models are alternative to Markov
Decision Processes, a largely used framework for this type of problems.
In particular, we showed that the state of a dynamical system can be represented more
compactly if it is described in terms of predictions of certain future events. We also showed
that even the cognitive process of selecting actions, known as policy, can be seen as a dynamical
system. Starting from this observation, we proposed a panoply of algorithms, all based on
predictive policy representations, in order to solve different problems of decision-making, such
as decentralized planning, reinforcement learning, or imitation learning.
We also analytically and empirically demonstrated that the proposed approaches lead to
a decrease in the computational complexity and an increase in the quality of the decisions,
compared to standard approaches for planning and learning under uncertainty.
