RL 002
· 2 min read
Sequential Decision Making
- A sequence is a series of events or actions that occur in a specific order over time.
- Goal: select actions that maximize total expected future reward.
- It may require balancing immediate and long-term rewards.
- It may be better to sacrifice short-term gains for long-term benefits.
- It may require strategic behavior to achieve high rewards.
Example of SDM
- Web Ads
- Agent AdBot
- Choose and Ad .
- Receives View time
- Receives Click on Ad
- Environment
- Receives the Chosen Ad .
- Provides the View time .
- Provides the Click on Ad .
- Agent AdBot
- Robot picking trash
- Agent Robot
- Moves arms to pickup trash .
- Receives Camera image of the room .
- Receives Reward if no trash on the floor
- Environment
- Receives robots action/arm movements .
- Provides Camera Image of the room .
- Provide reward of +1
- Agent Robot
- Robot making Pizza.
- Agent Robot
- Makes an action from Action space (a range of possible actions, Arm Movements, Get Pizz from Oven, Place Pizza on Tray, Cut Pizza...).
- Receives Camera image of the kitchen .
- Receives Reward if pizza is made correctly
- Environment
- Receives robots action/arm movements .
- Provides Camera Image of the kitchen .
- Provide reward appropriate to the action .
- Reward space
- +10 Successfully made pizza.
- +1 Get Pizza from oven.
- +1 Place Pizza on tray.
- +1 Get Packing box.
- +1 Cut Pizza successfully.
- Agent Robot
Rewards
- a scalar signal that indicates how well the agent is doing at time step . (agent's performance)
- Examples of rewards:
- Robot play soccor:
- reward for scoring a goal.
- reward for kicking the ball out of bounds.
- Chess
- reward for winning the game.
- Autonomous drone flying stunt:
- reward for following the intended trajectory.
- reward for crashing into an obstacle.
- Robot play soccor: