Multi-criteria Reinforcement Learning

By Konkoly Thege. 1998

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Towards Continual Reinforcement Learning: A Review and Perspectives (2020, background): “This setting can also be viewed with the lens of multi-objective optimization where potentially conflicting requirements make it important to consider concepts like Pareto Optimality (Gábor, Kalmár, and Szepesvári, 1998) for cases where it is not possible to learn an optimal policy for each task ”
Learning to Act by Predicting the Future (2016, background): “Vector-valued feedback has been considered in the context of multi-objective decision-making (G´abor et al., 1998; Roijers et al., 2013).”
Multiobjective Reinforcement Learning: A Comprehensive Overview (2015, methodology): “The actions are chosen based on the thresholds and a lexicographic ordering (the last objective is maximized at first) [12], [37].”
Stochastic dynamic programming with factored representations (2000, background): “, those that support some type of functional decomposition of the value function such as neural networks [6, 80] or additive structure [9, 31, 37, 47, 56, 72, 71]).”
Empirical evaluation methods for multiobjective reinforcement learning algorithms (2011, methodology): “Single-policy methods can use a diverse range of approaches to specify these preferences such as scalarising weights (Natarajan and Tadepalli 2005), objective thresholds (Gabor et al. 1998) and ranges in objective space (Mannor and Shimkin 2004), and it is important to establish which of these appro”
A Distributional View on Multi-Objective Policy Optimization (2020, background): “For instance, threshold lexicographic ordering approaches optimize a (single) objective while meeting specified threshold values on the other objectives (Gábor et al., 1998), optionally with slack (Wray et al.”
Multi-objective reinforcement learning using sets of pareto dominating policies (2014, methodology): “For instance, Gabor et al. (1998) employ a lexicographical ordering of the objectives, while Barrett and Narayanan (2008) define linear preferences on the different objectives.”
Linear Fitted-Q Iteration with Multiple Reward Functions (2012, background): “Gábor et al. (1998) demonstrated that an MDP with multiple reward signals can be well-defined and solved so long as we are given a fixed partial ordering on reward vectors.”
A Review of Safe Reinforcement Learning: Methods, Theories, and Applications (2024, unknown): “Although using a surrogate function to replace the cumulative constraint cost has appeared in [166], [190], [191], CPO [3] ﬁrst show their algorithm guarantees for near-constraint satisfaction.”
Urban Driving with Multi-Objective Deep Reinforcement Learning (2018, background): “[6] has a few drawbacks, especially in the DQN setting:”
Scalarized multi-objective reinforcement learning: Novel design techniques (2013, methodology): “For instance, [4] suggests a MORL method that uses a lexicographic ordering of the objectives and by placing minimal thresholds on certain objectives, policies are discovered that take into account these constraints.”
A Lyapunov-based Approach to Safe Reinforcement Learning (2018, methodology): “Another surrogate algorithm was proposed by [13], in which the algorithm ﬁrst computes a uniform super-martingale constraint value function surrogate w.r.t. all policies.”
A Review of Safe Reinforcement Learning: Methods, Theory and Applications (2022, unknown): “Constrained Policy Optimization Recently, Constrained Policy Optimization (CPO) [4] suggests computing the cost constraint using a surrogate cost function which evaluates the constraint J (πθ) according to the samples collected from the current policy πθk : Although using a surrogate function to rep”
Risk-Sensitive Reinforcement Learning Applied to Control under Constraints (2005, methodology): “Optimal policies generally depend on the starting state, and are nonstationary and randomized (Feinberg & Shwartz, 1999; Gabor et al., 1998; Geibel, 2001).”
A Survey of Multi-Objective Sequential Decision-Making (2013, methodology): “A widely cited approach to nonlinear scalarization is that of Gabor, Kalmar, and Szepesvari (1998), which is designed for tasks where constraints must be satisﬁed for some objectives.”