An Alternative Softmax Operator for Reinforcement Learning
summary
Introduces the Mellowmax (mellomax) operator: log of the average of e^{omega·Q}/omega (a softened Boltzmann-style average) and proves it satisfies a non-expansion contraction property so generalized value iteration has a unique fixed point and fast convergence. Experiments show Boltzmann sometimes fails (8 of 200 random MDPs nonterminating, 3 with multiple fixed points) while mellanax never does. Replacing max with mellomax in Deep Q-learning replaces max in DQN removes the target network and yields more stable learning and higher returns on several Atari-style games.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 5:05Pacman in a Grid
- 7:47NonExpansion Property
- 14:39Contraction Property
- 15:15Properties
- 16:13Algorithm
- 17:26Bellman Operator
- 18:54Boltzmann Operator
- 23:01Deep Q Learning
- 27:30Questions
related talks
57mComputationally Efficient Alternatives to Nonconvex-Nonconcave Min-Max Optimization
55mLearning Minimax Estimators Via Online Learning
31mQ-learning with Uniformly Bounded Variance
01h 00mMin-Max Optimization from a Dynamical Systems Viewpoint
55mOptimality and Approximation with Policy Gradient Methods in Markov Decision Processes
46mSmooth Nonconvex Min-Max Optimization
31mStable Reinforcement Learning with Unbounded State Space
35mThe Mean-Squared Error of Double Q-Learning
