Strong AI

When AI research first became popular it was driven by a desire to create an artificial intelligence that could perform as well, or better, than a human brain in every relevant aspect. May it be a question of intellectual capacity, a display of emotion or even compassion. This is what we refer to as Strong AI. It can more concisely be described as Can machines really think?, equating the epitomy of AI intelligence with human or superhuman intelligence.

The early view was that a strong AI had to actually be consicous as opposed to just acting like it was conscious. Today most AI researchers no longer make that distinction. A problem arising from holding the view that strong AI has to be conscious is describing what being conscious actually means. Many philosophers attribute human consciousness to human's inherent biological nature (e.g. John Searle), by asserting that human neurons are more than just part of a "machinery".

Some key words in relation to Strong AI are:

• Mind-body problem The question of whether the mind and the body are two distincs entities. The view that the mind and the body exist in separate "realms" is called a dualist view. The opposite is called monist (also referred to as physicalism), proponents of which assert that the mind and its mental states are nothing more than a special category of physical states.
• Functionalism is a theory stating that a mental state is just an intermediary state between input and output. This theory strongly opposes any glorification of human consciousness; asserting that any isomorphic systems(i.e. system which will produce similar outputs for a certain input) are for all intents and purposes equivalent and has the same mental states. Note that this not does not mean their "implementation" is equivalent.
• Biological naturalism is in many ways the opposing view of functionalism, created by John Searle. It asserts that the neurons of the human brain possesses certain (unspecified) features that generates consciousness. One of the main philosophical arguments for this view is the Chinese room; a system consisting of a human equipped with a chinese dictionary, who receives inputs in chinese, translates and performs the required actions to output outputs in chinese. The point here is that the human inside the room does not understand chinese, and as such the argument is that running the program does not necessarily induce understanding. A seemingly large challenge to biological naturalism is explaining how the features of the neurons came to be, as their inception seem unlikely in the face of evolution by natural selection.

Weak AI

A weak AI differs from a strong AI by not actually being intelligent, just pretending to be. In other words, we ignore all of the philosophical ramblings above, and simply request an agent that performs a given task as if it was intelligent.

Although weak AI seems like less of a stretch than strong AI, certain objections have been put forth to question whether one can really make an agent that acts fully intelligently:

• An agent can never do x, y, z. This argument postulates that there will always be some task that an agent can never simulate accurately. Typical examples are being kind, showing compassion and enjoying food. Most of these points are plainly untrue, and this argument can be considered debunked.
• An agent is a formal system, being limited by e.g. Gödel's incompleteness system. This argument states that since machines are formal systems, limited by certain theorems, they are forever inferior to human minds. We're not going to go deep into this, but the main counterarguments are:
  • Machines are not inifinite turing machines, and the theorems do not apply
  • It is not particularly relevant, as their limitations don't have any practical impact.
  • Humans are, given functionalism, also vulnerable to these arguments.
• The, sometimes subconscious, behaviour of humans are far too complex to captured by a set of logical rules. This seems like a strong argument, as a lot of human processing occurs subconsciously (ask any chess master – cough Asbjørn –for instance). However, it seems infeasible that this type of processing can not be encompassed by logical rules themselves. Saying something along the lines of: "When Asbjørn processes a chess position he doesn't always consciously evaluate the position, he merely sees the correct move", seems to simply be avoiding the question. The processing in question here might not occur in the consciousness, but some form of processing still happens somewhere.
• Finally, a problem with weak AI seems to be that there is currently no solid way of incorporating current knowledge and common background knowledge with decision making. In the field of neural networks, this point is something deserving future attention and research according to Norvig, and Russel.

Conditional probability tables

As you might have realised by now, using Bayesian Networks reduces the complexity of a model by assuming many variables independent of each other. But we still need to establish probability tables for all our variables, even if they don't depend on any other variable. We are thus going to extend our use of joint probability tables, and for the nodes in our system create conditional probability tables (often the acronym CPT is used). They are very alike joint probability tables, but for a given node, only the outcomes of its direct parents are taken into account. Thus, for the node/variable $Rain$, a CPT would look something like the table below:

$$ \begin{array}{|cc|c|} \textbf{Cloudy} & \textbf{Rain-dances} & \textbf{P(Rain = True)} \\ \hline True & Many & 0.8 \\ True & Few & 0.7 \\ False & Many & 0.05 \\ False & Few & 0.001 \\ \end{array} $$

Note that we omit the probability $P(Rain = False)$ as it is simply implied that it is $1 - P(Rain = True)$.

Transition models

In the scope of this text, we assume that a transition from one state to another is Markovian, that the probability distribution of results given an action is only dependent on the agent's current state, and not its history. You can think of the transition model as a big three-dimensional table for the function: $$P(s'|s,a)$$ Given the agent is in a state, $s$, and performs an action, $a$, it will reach a new state, $s'$ with the probability $P(s'|s,a)$.

Markov decision process

Assume you have a sequential decision problem for a fully observable, stochastic envirnment with a Markovian transition model. When the agent moves toward its goal, it is given a reward at each state it is in between its initial state and goal. This reward may either be positive or negative. As the agent moves, it aims to maximize its reward. This is called a Markov decision process, or MDP for short. Assume that at each step that is not a goal state, it is penalized with a negative reward of -0.1, and is rewarded +1 at the goal state. If it is able to reach its goal with 5 steps, it would have a utility of 0.5. If it used 10 steps that utility would be 0. A policy is a solution of a MDP. A policy is traditionally denoted $\pi$ and $\pi(s)$ is the recomended action for the agent when in the state s. An optimal policy is a policy that yields the highest expected utility and is denoted $\pi^*$.

A complete policy is a policy that whatever outcome of any action the agent takes, knows what to do next.

Utilities over time

When working with sequential decision problems we can roughly partition the problem into two main categories:

Finite Horizon

There is a fixed, finite time N after which nothing matters, the game is over. With a finite horizon, the optimal action in a given state can change over time. We say that the optimal policy, $\pi^*$, is nonstationary.

Infinite Horizon

There is no upper time limit, and consequently the optimal policy does not change with time, the policy is said to be stationary

Given that we have a stationary assumption, there are two ways that we can assign utilities to sequences:

Additive rewards

The utility of a sequence of states is $$ U_h([s_0, s_1, s_2, ..]) = R(s_0) + R(s_1) + R(s_2) + ... $$ Where $R(s)$ is the Reward function

Discounted rewards

The utility of a sequence of states is $$ U_h([s_0, s_1, s_2, ..]) = R(s_0) + \gamma R(s_1) + \gamma^2 R(s_2) + ... $$ Where $0<\gamma\leq 1$ is a discount factor. This makes sense when considering the increased uncertainty related to steps further away.

Optimal policies and utilities of states

Assume the agent is in state $s_0$ and and is executing the policy $\pi$. The subsequent states at time $t$ are described by the random variable $S_t$. Then the expected utility (when reaching a goal state) is given by $$ U^{\pi}(s_0) = E(\sum_{t=0}^\infty \gamma^t R(S_t))$$ Where the expectation is with respect to the probability distribution over state sequences determined by $s$ and $\pi$. A subset of these policies will have a equal and higher utility than the rest. These are optimal policies and such a policy is denoted by $\pi^*_s$: $$ \pi^{*} = \max_{\pi} U^\pi (s)$$

An interesting feature of sequences using discounted utilities with infinite horizons are that they are independent of initial state. This comes straight from the algoritmic property of shortest paths: If a path $a\to b$ is a shortest path between these two nodes, then a subpath $a' \to b'$ within $a\to b$ is also a shortest path.

The utility function $U(s)$ allows the agent to select actions by using the principle of maximum expected utility - that is, choose the action that maximizes the utility of the subsequent state $s'$:

$$\pi^*(s) = \max_{a \in A(s)} \sum_{s'} P(s'|s,a)U(s')$$

The Bellman equation for utilities

When in a state $s$, the utility of that state is the immediate reward of that state plus the discounted expected utility of the next state, given that the action chooses the optimal action, that is: $$U(s) = R(s) + \gamma \max_{a\in A(s)} \sum_{s'} P(s'|s,a)U(s') = R(s)+\gamma\cdot\pi^*(s)$$

This is called the Bellman equation.

The value iteration algorithm

Let $U_i(s)$ be the utility value for the state $s$ at the $i$th iteration. The iteration step, called the Bellman update is then: $$U_{i+1}(s)\leftarrow R(s) + \gamma\cdot\pi^*(s)$$

Filtering

Calculating the unobservable variables based on the evidence (the observable variables).

Prediction

Trying to predict how the variables will behave in the future based on the current evidence.

Smoothing

Estimate past states including evidence from later states.

Most likely explanation

Find the most likely set of states.

Building decision trees

The general procedure for building decision is a simple recursive algorithm, where you select an attribute to split the training data on, and then continue until all the data is classified/there are no more attributes to split on. How big the decision tree is depends on how you select which attribute you split on.

For an optimal solution, you want to split on the attribute which holds the most information, the attribute which will reduce the entropy the most.

1. Retrieval

Finding a case from the case base which is similar to our current situation or state.

2. Re-use

Retrieving a case and proposing it as a valid action to apply in our current state. May require som adaptation to fit our current situation.

3. Revision

Evaluating, through a series of simulations, how well the proposed action will perform and whether it's practical to apply in our case.

4. Retention

Storing the result of this experience in memory.

Euclidean Distance

$\sqrt{\sum_{i=1}^{k}(x_i-y_i)^2}$

Manhattan Distance

$\sum_{i=1}^{k}|x_i-y_i|$

Cosine Similarity

$sim(A,B)=cos(\theta)=\frac{A*B}{||A||||B||}$

Instance-based

Answer: Describe the four main steps of the case-based reasoning (CBR) cycle. What is the difference between “instance-based reasoning” and “typical CBR”?

An approach based on classical machine learning for classification tasks. Attribute-value pairs and a similarity metric. Focus on automated learning without intervention, requires large knowledge base.

Memory-based reasoning

As instance-based, only massively parallel(?), finds distance to every case in the database.

(Classic) Case-based reasoning

Motivated by psychological research. Uses background-knowledge to specialise systems. Is able to adapt similar cases on a larger scale than instance-based.

Analogy-based

Similar to case-based, but with emphasis on solving cross-domain cases, reusing knowledge from other domains.