How to prevent the eligibility trace in SARSA with lambda = 1 from exploding for state-action pairs that are visited a huge number of times?
I was testing SARSA with lambda = 1 with Windy Grid World and if the exploration causes the same state-action pair to be visited many times before reaching the goal, the eligibility trace gets incremented each time without any decay, therefore it explodes and causes everything to overflow. How can this be avoided?
If I've understood correctly your question, the problem is that the trace for a given state gets incremented too much. In this case, a potential solution is to use replacing traces instead of the classic incremental traces.
The idea in replacing traces is to reset the trace to a value (typically 1) each time the state is visited. The following figure illustrates the main difference between both kinds of traces:
You can find more information in the classical Sutton & Barto book Reinforcement Learning: An Introduction, especifically in Section 7.8.
- DQN model either doesn't work or it is extremely slow in training
- Deep reinforcement learning - how to deal with boundaries in action space
- QLearning and never-ending episodes
- Good implementations of reinforcement learning?
- Understanding policy and value functions reinforcement learning
- Best practices for exploration/exploitation in Reinforcement Learning
- exploration and exploitation in Q-learning
- iterations and reward in q-learning
- Reinforcement learning toy project
- Reinforcement Learning
- Reinforcement learning And POMDP
- Reinforcement learning with neural networks
- Reinforcement learning algorithms for continuous states, discrete actions
- Supervised learning v.s. offline (batch) reinforcement learning
- How to do backpropagation in PyTorch when training AlphaZero?
- Gymnasium custom environment "too many values to unpack" error
- How to train an artificial neural network to play Diablo 2 using visual input?
- PPO agent is not learning
- Pytorch Geometric graph batching not using DataLoader for Reinforcement learning
- Keras-rl2 error Compability with Tensorflow
- Monte Carlo Method for Blackjack: strange Q-values table
- Performance issue with gradient-bandit agent
- Probability 0 in Importance Sampling
- Complex state in re-inforcement learning
- OpenAi-Gym Discrete Space with negative values
- How to seed `gymnasium` environment resets when using `stable_baselines3`?
- Using Tensorflow Huber loss in Keras
- Training a Custom Feature Extractor in Stable Baselines3 Starting from Pre-trained Weights?
- AttributeError: module '_Box2D' has no attribute 'RAND_LIMIT_swigconstant'
- How do I log observations after reset in Stable_Baselines3?