Intelligent Agents and Environments

The agent view of AI

A useful way to frame AI is the agent: anything that perceives its environment through sensors and acts on it through actuators, choosing actions to maximise a performance measure. Environments differ: fully or partially observable, deterministic or stochastic, single or multi-agent, static or dynamic, discrete or continuous. Agent designs range from simple reflex agents (condition-action rules) to goal-based, utility-based and learning agents. The rest of AI is largely about building the decision-making part.

A reflex agent in a two-room vacuum world, run

I ran this with plain Python 3 (standard library only), with fixed random seeds where randomness is used. A simple rule (suck if dirty, otherwise move to the other room) cleans both rooms in three steps, then keeps moving back and forth because a reflex agent has no memory or goal to tell it to stop. Its score is 17 (10 per clean, minus 1 per move).

# A simple reflex agent in a two-room vacuum world
def agent(percept):
    room, dirty = percept
    if dirty: return "suck"
    return "right" if room == "A" else "left"
world = {"A": True, "B": True}; room = "A"; score = 0
for step in range(5):
    action = agent((room, world[room]))
    if action == "suck": world[room] = False; score += 10
    else: room = "B" if action == "right" else "A"; score -= 1
    print(f"step {step}: action {action:<5} -> now in {room}, dirty rooms {[r for r in world if world[r]]}")
print("performance score:", score)

Output:

step 0: action suck  -> now in A, dirty rooms ['B']
step 1: action right -> now in B, dirty rooms ['B']
step 2: action suck  -> now in B, dirty rooms []
step 3: action left  -> now in A, dirty rooms []
step 4: action right -> now in B, dirty rooms []
performance score: 17

Define the performance measure first

Decide what "doing well" means (clean floor, low energy, short time) before designing the agent; it changes the right behaviour.

Quick check: What does a simple reflex agent base its action on?

• Only the current percept, using condition-action rules
• A full model of the future
• A learned value function
• A conversation history