CSCE 420 Lecture 23

« previous | Thursday, April 11, 2013 | next »

Bayesian Networks

Each node represents an event that has a certain probability, and dependence / conditional probabilities are indicated by the incoming directed edges

"Lower down" nodes represent observable "symptoms", and they are connected through intermediate nodes to the original cause nodes.

1. Draw causal nodes first
2. Draw directed edges to effects ("direct causes")
3. Links encode conditional probability tables
4. Absence of link implies conditional independence

Advantage of networks: fewer params than JPT (for 5 boolean random variables, 2³⁵ = 32 entries)

Intelligent Agents

Characteristics:

• situated: in a dynamic environment (changes over time) in which agent can take actions ta change state of environment
• goal-oriented: actions appear to be driven to achieve some goal
• autonomy: independent of human interaction/input

Non-essential characteristics:

• adaptive: learn from experience
• "social": communicate/interact/cooperate with other agents (important in multi-agent systems)

Other vocabulary

• perception: input from the environment (through sensors)
• actuators (effectors): things that change the state of the world
• plan: action or sequence of actions to change states of world toward goal state
• policy: decision function of agent, expressed as $\pi :{\text{states}}\mapsto {\text{actions}}$
• KB/goals/model: what the agent is "thinking"
• utility: A score of "goodness" or "desireability"; an alternative way of representing goals ($u:{\text{states}}\mapsto \mathbb {R}$)
• performance criteria: Our measure of the agent's actions
• rationality: "doing the right thing" taking actions to maximize utlity given what agent knows

Environments

1. Discrete (symbolic representations) vs. Continuous (control theory)
2. Static vs. Dynamic: whether the state of the world changes while agent is making decisions
3. Deterministic vs. Stochastic
   • outcomes of actions, probability distribution over outcomes (pulling trigger may not fire bullet)
   • other actors, uncontrolled events
4. Episodic vs. Sequential: can't necessarily make short-term decisions
5. Fully Observable vs. Partially Observable: does agent have access to full state of the world?

In AI, we generally stick to discrete, static environments