26
X-Machines for Agent-Based Modeling: FLAME Perspectives
• Reactive units. Using evolution or Darwinian system, agents react
to changes in system and adapt their behavior.
• Goal-directed units. A few agents will be working towards achieving their goals, such as companies taking over other companies for
growth or power.
• Planner units. Agents who are goal-directed but also consider environment and goods in their strategy planning.
Agents can use built-in tools to perform parameter learning and assess
their actions as they behave in the simulation. This can be achieved
through supervised, unsupervised or reinforcement learning, depending
on scenarios being modeled. This allows agents to change their strategies
or functions depending on personal preferences and information received.
Adaptive agents use multiple methods to learn about the system. Holland and Miller [90] use genetic algorithms to model a population of
solutions, coded as strings of characters. Genetic algorithms learn by a
biased search towards a combination of solutions, using crossover and
mutation. Other methods like classifier systems are an adaptive rulebased system, where each rule is in condition-action (if-then form). The
condition allows specific actions to take place.
Reinforcement learning determines how agents can maximize their goals.
This differs from supervised learning by finding a balance between exploration and exploitation. The state of the agent at any given time s t is
chosen from a set of allowable states S. The state also determines which
action will be chosen A(s t ).
s t ∈ S choose action a ∈ A(s t )
An agent finds a policy π : S → A to maximize its reward r = r1+r2+r3.
Various kinds of learning include role learning, learning by discovery and
observation through experiments.
Each method is tailored for the problem modeled. Using evolutionary
techniques allows agents to make independent decisions because
Agents are autonomous. Agents can operate without intervention of
other agents.
Agents are reactive. Agents can read the environment and other
agents actions to react accordingly.
Agents are proactive where each agent works to satisfy a specific
goal.
Agents are social where they interact with other agents through communication frameworks and alter behavior accordingly.
X-Machines for Agent-Based Modeling: FLAME Perspectives
• Reactive units. Using evolution or Darwinian system, agents react
to changes in system and adapt their behavior.
• Goal-directed units. A few agents will be working towards achieving their goals, such as companies taking over other companies for
growth or power.
• Planner units. Agents who are goal-directed but also consider environment and goods in their strategy planning.
Agents can use built-in tools to perform parameter learning and assess
their actions as they behave in the simulation. This can be achieved
through supervised, unsupervised or reinforcement learning, depending
on scenarios being modeled. This allows agents to change their strategies
or functions depending on personal preferences and information received.
Adaptive agents use multiple methods to learn about the system. Holland and Miller [90] use genetic algorithms to model a population of
solutions, coded as strings of characters. Genetic algorithms learn by a
biased search towards a combination of solutions, using crossover and
mutation. Other methods like classifier systems are an adaptive rulebased system, where each rule is in condition-action (if-then form). The
condition allows specific actions to take place.
Reinforcement learning determines how agents can maximize their goals.
This differs from supervised learning by finding a balance between exploration and exploitation. The state of the agent at any given time s t is
chosen from a set of allowable states S. The state also determines which
action will be chosen A(s t ).
s t ∈ S choose action a ∈ A(s t )
An agent finds a policy π : S → A to maximize its reward r = r1+r2+r3.
Various kinds of learning include role learning, learning by discovery and
observation through experiments.
Each method is tailored for the problem modeled. Using evolutionary
techniques allows agents to make independent decisions because
Agents are autonomous. Agents can operate without intervention of
other agents.
Agents are reactive. Agents can read the environment and other
agents actions to react accordingly.
Agents are proactive where each agent works to satisfy a specific
goal.
Agents are social where they interact with other agents through communication frameworks and alter behavior accordingly.