Day 8 - Exploration vs Exploitation
09 May 2023Morning Session: Exploration vs Exploitation
- Zibra fish:
- Different populations of neurons coding for different states
- They keep on oscillating between the states: in one instance: they look for food (they stay locally), in another instance they move around (exploratory behavior)
- By looking at the neural activity (Ca2+ imaging), we can decode which state they are currently in but also predict for how long they would stay in this state: we can see a decay in those state-populations
-
“Constraints make things interesting”
- Strategies for exploration:
- Aside from external reward, in cases where those are sparse, the agent can have internal rewards (rewarding itself, intermediate ones)
- RL: peter Dyan:
- Take reward function and replace it with add additional term (bonus: novelty)