2.2 KiB
AI Agents
AI agents are defined as a program (state of all previous actions and inputs) that produces an agent function
Reflex Vacuum Cleaner Agent
Essentially a rules table, based on the state of the environment and the agent, (state of dirt in a/b, position of cleaner in a or b)
Rational Agents
A fixed performance measure the agent should set as their goal, based on the environment and the agents actions
- One point per square cleaned in time T?
- One point per clean square per time step, minus one per move?
- Penalize for > k dirty moves? A rational agent chooses the action that maximizes the expected value of the performance measure given the percept sequence to date.
A rational agent is not omniscient or clairvoyant. It is not guaranteed they will perceive everything relevant, or are given the relevant information. They also may not be successful in the action as expected. A rational agent does not mean a successful one.
Defining internet agents
Automated Taxi
- Performance measure
- environment
- actuators
- sensors
Internet Shopping Agent
- Performance measure
- environment
- actuators
- Sensors
Task environment characteristics
- Fully/Partially observable - Agents sensors give it access to complete state of the environment at each point in time
- Single/multi-agent - Single agent solving a problem (puzzle vs chess)
- Deterministic/Non-deterministic - Whatever next state is determined completely by the current state, ie randomness
- Episodic/sequential - Agent's experience divided into atomic episode (agent receives a percept and perform a single action)
- Dynamic/static - If environment changes while an agent is deliberating
- Discrete/Continuous - State of the environment (states in chess vs taxi driving)
- Known/Unknown - In known environments, the outcomes for all given actions are given
Agent types
- Simple reflex agents
- Model based reflex agents with state
- Goal-based agents
- Utility based agents All of these can be turned into learning types
A learning agent has 4 main components, a performance algorithm, a critic, a learning element, and a problem generator, which forms a feedback loop which enables the agent to learn to adapt to it's environment.